The result is decent, but it had a few bizzare/trivial syntax errors I had to fix manually: it would do an extra closing bracket or paren a few times, and wanted to separate function definitions with comma. Not sure what that was about, but otherwise the output run just fine.
So, with those qualifiers, I think it's a decent local coding model. It roughly compares with GPT-4.1 (!!), released 14 months ago, on the output: https://senko.net/vibecode-bench/2025/minesweeper-gpt-4.1.ht... (actually I'd call it better, but those syntax errors...)
I ran the quantized version (4-bit GGUF) on my consumer-grade card with 12G of VRAM and got 5t/s for output. Not for interactive use for coding, but fairly capable model.
To me, it's fascinating how much progress we got in over a year. GPT-4.1 was considered an extremely capable coding model. Now we got something with 12B of params performing roughly the same (in this specific benchmark, disclaimers, etc).
It was almost certainly not trained for coding, as it's got both audio and vision input, is only 12B, and nowhere in the announcement is coding mentioned. It will likely not have good performance on coding in general, compared to other small models like Qwen 3.6 35B A3B, Gemma 4 26B A4B, Nvidia Nemotron 3 Nano 30B-A3B, gpt-oss-20b.
For 16GB laptops, Qwen 3.5 9B is the undisputed champ.
Gemma 4 31B is the top dog at small model coding, but is dense so it needs ~48GB unified RAM for full context. If you want decent coding on a laptop you need a lot of RAM. But this shouldn't be surprising, dev machines have always needed lots of resources.
>consumer-grade card with 12G of VRAM and got 5t/s
That speed for token output indicates to me that it somehow is using hybrid mode and involving cpu+system ram somehow. That ~5tk/s is about the ram bandwidth of DDR4 RAM versus that size model at 4bit. Any consumer GPU with 12 GB like a nvidia rtx 2080 or rtx 3060 should be doing 20+ tk/s with llama.cpp and CUDA backend.
> It roughly compares with GPT-4.1 (!!), released 14 months ago
I think the mayor win for coding was reasoning. That's why such a small model can match GPT-4.1 in coding, but I suspect that GPT-4.1 still wins in general world knowledge due to bigger size.
> I suspect ... still wins in general world knowledge due to bigger size
Encyclopedic knowledge matters relatively little in perspective, given the expectable future developments: even the more knowledgeable of us will use that knowledge for reasoning and intuition (and we will have absorbed the intellectual keys during our training), but under our professional hat we should in theory be ready to go "I stand corrected" and "more precisely" with the actual data at hand.
I.e.: for the encyclopedic knowledge needed, the /understander/ will have a RAG subsystem and a corpus of knowledge to inquire upon processing queries.
(Corroboration: we can't delirate, and neither can the machine...)
Thank you for sharing this. Do you think the syntactical issues could be addressed with fine tuning or some other kind of parameter tweaking? That's frustrating hah.
With a harness you could feed the code to a linter and if there are errors feed that to a model automatically. It’s amazing that the models are good enough that I haven’t bothered doing this
The big story here is the encoder-free part, which I still don't fully understand.
> Vision: We replaced Gemma 4’s vision encoder with a lightweight embedding module consisting of a single matrix multiplication, positional embedding and normalizations.
I dont see how encoder free audio isnt a mistake here. a mimo model will at least get the audio to 12.5 Hz as opposed to the 25 Hz they are doing. and you dont need to finetune mimo either.
I've been waiting for something like this to be released since then.
The annoying thing is that chameleon was multi-modal out based on the same principles, but this model is just inputs... (I'm curious how they did pre-training without having multi-modal outputs as well. I wonder if they just chopped them off rather than support image output).
Some of the FAIR people moved to Thinky, and they also started doing encoder-free MM-LLMs. Now Google. This seems to becoming a trend working at small scale, but the difficult part is scaling.
Standard approach for training MM-LLMs is we train the encoder first, there are O(2-10B) good images on the internet, so encoder needs to see each image O(10-100) times, that is O(100T) tokens, which is more than the entire pre-training budget for most runs. That is the reason we train the encoder separately (smaller model, 2B active vs 30B or 200B active LLM); there is nothing magical about training the encoder and LLM together, it is just more token-efficient to train the image modality first.
Vision embedder (35M parameters): Replaces the 27 vision transformer layers of the other medium-sized Gemma 4 models. Raw 48x48 pixel patches are projected to the LLM hidden dimension with a single matmul. A factorized coordinate lookup (X and Y matrices) attaches spatial location information directly to the input
the "single matmul" is the key here, I haven't tried it, but it's probably pretty fast and memory efficient.
Not at all, I had the same feeling as yours the first time I read it. I think the key is that the "encoder" they're using is just a linear projection, which is probably pretty fast and memory efficient. A single matmul vs a ViT encoder is probably a huge win.
Not at all. Getting really pedantic, tokenization is also a form of encoding, so it doesn't matter the modality you're using, you'll end up doing some type of encoding in some way.
Tokens are such a strange base unit. Couldn't we do something that naturally conforms better to reality than such choppy units that cause all sorts of artifacts? making everything 'language based' prevents true multi-modality. Thinking isn't done in language. Thinking outputs language, but its far more like multiple waves of data coalescing into an 'idea', internal... subjectively (n=1) at least. I think wave/signal based transformers are the next jump.
After that a s1/s2 system: fast generation, slow wave correction / observation operating over the fast generation seems like the next leap forward.
Tokens create and hide too many problems to be the 'optimal' solution.
Not to be too snarky but there’s a few trillion dollars and some of the brightest minds of our generation working on this. I’m sure there’s a reason why they’ve settled for or are stuck on tokenization.
I agree with the GP. The idea that there's not a better intermediate representation between tokens and embedding vectors seems absurd. But how to arrive at such a representation and implement it effectively is a few zeroes above my pay grade.
I find your agreement seductive because it side steps the unfounded assertions and simply asserts there must be something different and we don’t know it, which is easy for me to agree with too. Or maybe hard to disagree with.
Smaller models are less forgiving to quantization. For a 12B model I wouldn't expect Q4 to be "pretty close", unless it underwent quantization aware training (QAT). Of course it's not set in stone, there's a huge variance between models, so this might surprise.
The audio side is even more interesting, as it seems they totally got rid of positional embedding are just doing a single linear transform to match the LLM input dimension and that's it.
> Audio: We simplified audio processing even further. We removed the audio encoder entirely and projected the raw audio signal into the same dimensional space as text tokens.
I guarantee you there's positional information one way or another. they just don't mention it because positional embeddings are extremely cheap computationally, not worth mentioning
Ah yeah, thinking further it's probably just using some positioning embedding based on sequence numbering added in the LLM layers. For vision it needs the patch location as well.
From the visual guide, there's still the 35M parameter embedder, then the linear projector, for vision, and the linear projector for audio, so it does have some parameters used for the multimodal input to project it into the LLM latent space: https://newsletter.maartengrootendorst.com/p/a-visual-guide-...
And the Unsloth quants, which are missing this, don't support multimodal input. (edit: actually, I may have just needed to update my llama.cpp, will check with an updated llama.cpp soon)
I'm downloading the ggml-org GGUFs now, I tried Unsloth but got some weird problems, double checking with the bf16 model to see if the issue was just the quant.
Anyone with a 16GB Mac — that is quite a lot of journalists, surely — can download that, install a model into it, and play.
Surely journalists have to start asking questions at least about OpenAI's consumer revenue projections now.
I am a major, major AI cynic, but I decided to be an informed cynic so I've been playing with local models for agentic work and a bit of CAD-to-image generation. I really quite like the 26B Gemma model — I've been using it to teach myself some fundamental things and learn OpenCode without developing a cloud dependency. It writes fairly good code and it is helping me learn the things I want to learn at a pace that I prefer.
But if this 12B model is even half as close as they say it is, this casts some doubt on the consumer end of the cloud business model, at least in the short term.
(Not clear if this app is using the MTP drafters; I've still not got them working with Gemma myself, though the Qwen 3.6 built-in MTP support is super in LM Studio)
I had discounted Edge Gallery because it didn't support system prompts, but now it does so I will give it another go. I believe the implementation does use MTP since I got an update to Gemma-4-E4B on iOS indicating such, and on macOS it's very speedy.
However, on my 18GB RAM MacBook Pro, selecting Gemma-4-12B-it results in this error:
> The model "Gemma-4-12B-it' requires more memory (RAM) than is available on your device.
So yeah, my questions about the 16GB marketing copy are fair.
Interesting; they may have fluffed up somewhere then.
(Though perhaps it'll squeeze in with a small context window? Not sure I understand that aspect yet)
It does seem to use MTP, yes, and it is quite quick — seemingly the underlying LiteRT stuff can do MTP with Gemma 4 and presumably MTP is a big part of the practicality picture here.
The system prompt thing was a surprise when I poked around.
I don't think we've bottomed out on what we can do with embedding models. They're these tiny models that absolutely rip on modern cpus with 8 bit int optimizations. Like in my app we can say pretty definitive things about hundreds of millions of places in the world on retrieval tasks on regular hardware.
Either Google changed the text or you editorialised it a tiny bit - just for all others that got excited, they mean 16GB VRAM. So a premium graphics card requiring a >2500€ device is the minimum to run this.
Still progress, but not quite democratic yet.
Weird though that Google might be cannibalising it's own AI subscription service?
I haven't tried this model yet, but I can run Gemma 31B w/ the MTP drafter in pure CPU at about 10tok/s so this should run at about 20-30tok/s on a decent CPU, it'll probably run at >50tok/s on any Mac that can fit it, and lots of people have a gaming GPU with enough VRAM. In terms of access to hardware being a gate, it's one you can hop pretty easily.
Could you outline how you are running the MTP drafters? I've tried LM Studio but no dice there. I'm probably missing something but I think llama.cpp and Ollama can't do it yet either?
I don't have the energy to build stuff all the time, that's a rabbit-hole side tunnel I don't really want to get into. I have larger concerns in my life that are more urgent than developing that side of things.
But I've moved on from Ollama for the time being, though I am mainly interested to see what the Gemma 4 MTP speeds are like on my M1 Max, so I may test it.
I am quite impressed with the tools in LM Studio, which is also a beautiful app, but it is not open source (which challenges my personal strategy somewhat) and I dread its inevitable enshittification.
Nevertheless the GUI has been very helpful while I learn, and I will probably use it until something else presents or my usage pattern settles down from experimentation to something a bit more routine.
I will try oMLX, too, but judging by the LiteRT page I may soon be able to just use that for the larger models if I end up settling with Gemma 4.
I haven't yet pushed the MTP enabled gemma4 12b model for Ollama because in my testing I wasn't getting a performance bump. The other gemma4 MTP models should work OK right now, but there are some fixes we're just about to push. This is specifically for the MLX backend.
Thanks for your reply. I will go back and look at Ollama again.
So much to learn but this news has really vindicated my decision to direct my limited span of concentration and focus to learning how to use open weights models and opencode.
I think the idea is that the model is seeing embeddings that map directly to underlying pixel data, rather than being fed semantically rich embeddings from an encoder model which itself had seen the raw pixel data.
Encoder free is huge for running on SBCs etc. often the encoding time is a significant fraction of generation time if you are using a VLM as a all purpose vision model
It actually works well because unlike encoders, the latent space is trained on that initial layer so it “knows” what to do with that sparse density. I’ve been using gemma4-12b with Flux2 and its ability to reason on visual input is pretty good. That said, each model is good in their own ways so YMMV but overall, it’s about as solid as Qwen just with a more advanced architecture.
The guide describes it as projection although there is apparently an extra step: "A factorized coordinate lookup (X and Y matrices) attaches spatial location information directly to the input."
12B at int8 would take up 12G memory, or 75% of the system memory which technically fits within 16GB but the OS will not like that. EDIT: On my 18G memory MacBook Pro, LM Studio reports a "partial GPU offload" for the int8 MLX weights. Can't test because the `gemma_unified" architecture is NYI.
We are now entering the closed loop game. Google doesn't need anyone else to accelerate their models. This is their bread and butter.
I'm both shocked but also not surprised that they continue to develop such efficiencies. Honestly it's like silicon and CPU architecture advancement. We kept shrinking it and shrinking it and it kept getting more and more powerful and here we are with AI and it's only going to be 100x more efficient with time. Maybe there's some point of decay but essentially the next 30 years will be more advanced than the last 30 and were going to be living in some sort of futurist blade runner scenario where gene editing is repairing ageing cells, organs and curing all sorts of cancers that haven't even appeared yet. Beyond our lifetimes people will live to 125 quite steadily and with great mobility and then obviously people will look to how do we get to living 1000 years, which of anyone is religious knows Noah and others lived to that age in a totally different era.
Anyway I'm going off on some tangent but look back 30 years. Now look forward 30 years. It's going to be insane. May God protect us.
> We kept shrinking it and shrinking it and it kept getting more and more powerful and here we are with AI and it's only going to be 100x more efficient with time.
It's definitely an exciting time, but in terms of advancements in the state of the art, there is a lot of low-hanging fruit left to pick. There IS a bottom, however, as you can only encode so much "knowledge" in a small number of parameters.
This feels to me a lot like what the early days of what radio or aviation must have been like. Or, heck, microcomputers even.
It's definitely a core component of a bigger system. We are effectively trying to recreate intelligence and human life through models and robotics. So the key insights for me, the LLM is the cerebral cortex but we have a lot more to recreate. Once you map in sensory input continuously and give it physical robotics, things start to change. But even before that leaving these things in simulated realities is what will happen, and right now we have things that operate based on our commands, but a complete step function will be the things that act on their own and that will be a very dangerous time but also where we see some very surreal things happening. They might not necessarily be made in the same way either, they might operate on entirely different types of architecture.
Yes I've taken the "must optimise longevity" route, taking priority over other things such as my career and hobbies. I want to see the future - all this AI stuff fascinates me.
What's Google's business case for releasing open models? Don't get me wrong, I am grateful and appreciative of these releases. I'm trying to understand how it fits into their bigger picture as a for profit company? Are they not helping competitors build on the novel technology they have developed?
Is it simply goodwill and/or marketing? Or am I missing something strategic?
A big part of the frontier labs abilities to charge 80% gross margins on inference is having the cornered resource of frontier models.
If that inference becomes popular and valuable enough that those companies make billions of dollars in profit, those companies could use that profit to fund the building of alternative products and platforms that dis-intermediate google's relationship with the customer.
Google already has an 80% gross margin business, the biggest one in the world. Everybody wants a slice of it.
By offering frontier inference closer to cost and open-sourcing everything that's sub-frontier, they're commoditizing frontier labs' models, which inhibits their ability to durably make high gross margins on inference.
A 12B-sized model is a far cry from "frontier inference". That's more like DeepSeek V4 Pro territory which is a 1.6T model. Or for multi-modal models, Kimi 2.6 which is 1T.
2. I'm not sure what this has to do with the case, unless you're arguing Google has an ads monopoly, in which case the best argument would likely not be that adwords lead to bidding wars because that just sounds like they're selling a product people really want to pay for
3. There's nothing criminal about being a very diversified business
You're right that it's not literally frontier. But like recent Qwen releases, it is a lot more capable than anybody thought models of this size could be a year ago, like capable enough to set a ceiling on what you can charge for AI for certain applications. Others still clearly justify a stronger model, but this trend may continue, etc.
This won't replace commercially viable, revenue generating alternatives of their own devising, but it does enable development activity and initiate conversations with enterprises who start with this model but want to do slightly more.
That's my experience right now... my company is all in on a plethora of platform products. Also, Microsoft just yesterday said their goal was "Unmetered intelligence". There's a lot of things that can be enabled by small local models, and those things are part of stacks that can generate revenue in other layers.
Could say the same for camera processing in the Pixel Camera app or any other binary someone wants to re-use that comes included in a software distribution (seemingly for 'free'). They can't lock the instructions up on the server so they might as well make the binary be freely distributable?
Companies don't commonly give away executable binaries "just because", why'd they start now for these binary blobs that are the models?
Not that I'm unhappy about it! Yay for open data any day, I'm just not understanding why, at least beyond PR in nerd circles
Binaries are source code outputs, they are copyrightable and patentable. Weights are not copyrightable so people can freely extract the weights and run them. If Google patents any of the novel algorithms here releasing it all freely isn't an impediment to making people license it.
Because a model like this can't be as easily obfuscated as image processing. Image processing is a bundle of many moving parts, a lot of functions each with it's own inputs and outputs. A model is a single function which can be easily extracted and reused, in comparison
They could lock them down legally which would prevent commercial use, but they choose not to, and they boast about how many tens of millions of times Gemma models have been downloaded by developers.
So there must be more to the rationale than just local model weights getting hacked out of devices.
Google is one of the few verticalized options in AI: Data, models, cloud services, low-level silicon (TPUs), internal use cases, retail use cases, B2B uses, distribution (browser & mobile), etc.
They rise with the tide of AI adoption. But they gain ground if people opt into Google solutions. And any token sent to a Google model (free or paid) actively punishes their competitors that are then required to spend vast sums to remain bleeding edge.
If you're an AI lab, you definitely want research teams in this space - as this is where you can most easily iterate and make improvements which you'll then bake into larger, frontier models.
The question is: do you want to release your models, or use them purely for R&D?
Since everyone else is already releasing models of similar qualities, it's hard to say you're shooting yourself in the foot if you join the chorus.
The added cannibalization of releasing them is effectively zero, so the reputational benefits are likely to be worth it.
>The added cannibalization of releasing them is effectively zero, so the reputational benefits are likely to be worth it.
Nobody would be looking at Qwen if their ~30b class models weren't fantastically good, it's great advertising and builds significant goodwill with developers, who are going to be your biggest advocates.
The other thing is, all these models are already disposable grade, and in a year they'll all be outclassed by The Next Big Thing. "Open" models are less than 18 months behind SOTA right now and I can't imagine that will slow down much over the next two years, they may even begin to close the gap. Nobody even talks about llama 4 anymore despite only being a year old.
Neutering OpenAI and Anthropic would be my guess. Commoditized LLMs won't hurt Google nearly as much as it hurts the LLM-only companies, and so accelerating the inevitable just helps knock out potential future competition in areas where Google -does- make a lot of money now.
I sincerely wonder why. Chinese censorship is only really relevant if you're doing anti China stuff, which is to say never, while the Western kind of model censorship ( a combination of copyrights and general fairness ) are something everyone's had to work around at least once, even if just for writing an interesting story.
As long as Chinese firms are releasing good open models I imagine there isn't a huge downside for Google to release state of the art small models to compete in the "free" space.
It's to destroy possible footholds for competitors and prevent them from making money in segments that Google doesn't care too much about, but can trivially commoditize.
A strong business case for Gemma includes fine tuning, adding AI to apps that run in the cloud, strengthening Android, shifting unprofitable small AI compute to devices, and harming competitors. The first two would be done using Google's cloud services due to integration with Gemma. I think Google is currently the best positioned company to profit from AI sales to businesses over the next few years, and Gemma is a critical part of the story.
I think its even more puzzling because you can't even run Gemma 31b on google cloud, they only let you test it with a rate limit. No way (I can find) to actually pay them to use it.
We saw great results in our usecase using google direct. Moved to Openrouter because google wouldn't let us use it beyond a test.
Then Openrouters performance looked worse, not sure if there was a quantized version or something. So we instead looked at Deepseek v4 Flash, and opted to go for that.
This model would probably be great for a super low cost cloud model, would love to use it in the cloud, Google makes you go elsewhere.
Google's MO since always has been to release great products or services for free, position themselves high and then abandon them or just find uses for Enterprise sales.
I'm pretty sure they are doing it because they get some research experience by shrinking and improving these models, and because they know that by doing this they get some good PR among the dev community.
Google's "free" is and was ad-supported, even if some products now have a paid tier. These models don't include ads. Doesn't seem like the same underlying reason
Maybe they are hedging against a future where local models are just as good as cloud models? Or maybe they can go the Taalas route and start hardcoding Gemma on a chip and hardware manufacturers can use it for local private AI.
The complete Chinese worldwide domination in this sector would be the alternative, since nobody else is releasing anything meaningful.
Plus every open model undermines their local competition by furthering open research and reduces moats, especially since Gemini as a frontier model isn't really competitive with GPT nor Claude for most applications.
Isn't Apple about to license some variation of this from google for on-device AI? Maybe it’s their sales pitch to Apple and then they will lock it down.
Competition from Chinese alternatives hopefully forces more openness and efficient models. DeepSeek for example is nearly on par and far more resource efficient, good for the planet imo
Its image processing is terrible. I ran several tests against it against Qwen 3.5 0.8b (yes, 7% the size) and Qwen beat it every time with Gemma often getting things entirely wrong. I even gave it a plain image saying "This is a test" and it thought for 6 minutes trying to analyze it and failed. Qwen 3.5 0.8b confidently got it in under a second.
It may be that the Q6 quant I got is borked (or my LM Studio is), but either way, the 0.8b's performance is mind boggling in comparison.
I guess Google implements more / stronger guard rails than Alibaba and thus confuses these small models. At least this was my impression with Gemma3 models where it often said that the image contains some nudity / sex scenes and therefore it cannot give a description of the image. Never understood the point of this behavior....
The biggest problem with all the Google models has always been RLHF, particularly safety training. They take a good, smart model and make it behave like a corporate person that has been to far to many forced anti-{sexism, racism...} seminars so that it is now living in fear of saying something that could be construed as wrong by some moral standard.
The Qwen series adopted vision wayyy earlier than anyone else. No idea why the other labs were sleeping on it but they had about 2 years of experimentation without any competition.
Quite aside from the architectural changes, I suppose this is the answer to why Google had such a glaring hole in the (pretrained) Gemma4 model lineup between the Gemma4 4b and Gemma4 26b models!
A model that comfortably fits in 16GB of VRAM (allowing room for context) is a welcome upgrade.
I have vLLM running on a Linux machine in my basement, connected with Tailscale, and I use small models as part of tasks like this:
- Transcribing scanned documents into formatted text
- Captioning/describing images and classifying them for audience suitability (includes anti-spam)
- Matching documents with relevant Wikipedia pages for tagging
I don't use them like frontier models. I break the work down into micro-tasks with one clear goal for each prompt. I write a lot of glue software to make the complete flow work. I was working on all of these tasks before LLMs appeared on the scene. The LLMs have allowed me to replace a lot of complicated code with less code plus a model, while achieving better results.
I use local models for reasons of cost and control. I already had the workstation and GPU. The only running cost is electricity. I have used proprietary models from OpenAI and Google for some of these tasks, but I also encountered churn when the models I built my tools around were retired. I don't worry about that when I have the weights saved locally.
I've got a home-built dictation app that uses a local model to clear up the text and fix grammar. It was super easy to build. I’m extending it to capture meeting notes and summarise too. All on-device.
I saw a little app the other day, I think someone posted on here, that looks at your screenshot and renames the file based off the contents of the file.
There's tons of little examples like that. For a lot of use cases, you really don't need the frontier models.
I think small models have a very good niche for specific tasks. I utilise a fine tuned Phi-4 model (smaller than this one) that fits in about 3.5gb of RAM (not vram) for the document processing side of things for the desktop app I develop (a bit of a shameless plug - whistle-enterprise.com).
If you have a very specific idea for local model use you can find a way to make it work very well, you don't even need to have a graphics card or NPU chip. You just have to be extremely constrained in how it's used. I think as a generic chatbot they're not great, I'd use a hosted SOTA model and I'm a big fan of local LLMs myself.
Yea absolutely, but man, where to even start, it is very specific.
Fundementally I didn't use any wrappers like unsloth or axolotl, although I have used the latter before a year or two back and it was good, but I needed something very very custom. I also wanted the whole fine tuning pipeline to exported OpenVino model to be seamless.
I heavily leaned on codex, claude and some manual sleuthing around the internet to understand what I needed. I'd played about with QLoRA finetuning with axolotl before and felt most comfortable with that. So I needed to keep everything as stripped down as possible and figured I can just utilise the 3 main huggingface libraries (transformers, peft and datasets) and also bitsandbytes (as suggested by claude to quantize the model to keep this working on my GPU) along with some custom scripts generated by claude/codex (each cross referencing each other) that will do the different stages of the training run.
The next part was the data. Obviously didn't have access to thousands of meetings and associated output documents but I did have a 3090ti sitting there and a codex subscription. So I set about working out what format I needed the data in (many thanks again, to claude/codex) and started generating hundreds of different transcripts, different amounts of speakers, content, tones, subjects, spelling mistakes - like all the different things you could think a meeting would have. Then it's a case of actually generating a good meeting document off the back of the transcripts and creating the "gold standard" that we'd use.
I'm going to gloss over a lot here as I'd rather not detail it as it relates to some propriatary stuff that I had to work through, but you basically pair the transcripts together and run the training.
At the verification stage, there was pretty much 3 things:
1. "just" do some regex string matching to see if there's any of the source transcript key facts in the output to ensure fact preservation. Same with owner fabrication (who said what), I don't want something attributed to someone when it wasn't them that said it and then finally markdown validation.
2. Using codex/claude to validate the transcript and output from the model - I used the latest frontier models, probably overkill for my task, but they were good at the job
3. Finally me going through some actual recordings of myself, groups, meetings and manually verifiying the output
So a fair bit of work, and for context I'm on version 10 now, so it's been a journey!
I don't know about this model, but the next one up, the 31B I've been using as an agentic coding assistant in OpenCode, and basically anything that's easy enough that I'd trust Sonnet to handle, I trust Gemma 4 to handle and it's been doing a great job, it surprises me positively much more often than negatively. I not infrequently run into situations where Gemma 4 fails to do the task and I switch to Opus 4.7 and it fails also.
I use small models like Gemma to improve transcriptions from ASR models amongst other micro-tasks. I actually built out a fine-tuning whisper pipeline with all local (smaller) models meaning no cloud/big-tech co is able to train/sell my (private) data.
I use them for research on new features. If my feature is going to interact with a frontier language model in prod, I start with these free local ones which are all competent enough to produce structured output, make tool calls, interact with mcp etc. I don’t care much for the content at the early phase of engineering, I care about the schema & failure modes.
Then when I’m getting close to feature-complete, I’ll move to a hosted frontier model for the final integration.
Cost savings are enormous if you’re making dozens of calls to language models a minute.
"Small" models are the ones I can run myself on my own terms. LLMs aren't useful enough for me to justify spending hundreds of euros on a GPU with 16GB VRAM or something, and that's assuming I have the rest of the desktop just laying around. Back when I checked (before the RAM price hike), these models weren't meaningfully better than 4-8GB ones anyway, you'd have to go for the top tier cards at 24 or 32 GB iirc to get something vaguely in the direction of the SaaS versions, and that was absolutely out of my budget. Even if that changed, so have hardware prices so it'd probably still work out the same
I do not know which model specifically, but I saw the founder answering a question about how it's a small model that's focused on just this one specific requirement.
Quickly deployed it to check some benchmarks relevant for German language. These are results for CohereLabs/include-base-44 german only : Gemma 4 12B %61.9
The quwen 3 14B vs Gemma 4 12B difference is within random variance they same in some repeat runs they actually got the exact same score. Next step up Gemma 4 31B gets 0.676 on this. Or let in some reasoning Qwen 3 14B (reasoning) 0.676.
I'll run some cheat-proof benchmarks ones tomorrow see if qwen is still on top.
This is a pretty good update. The demo video is a bit funny though - the tester asks to turn the release into bullet points. okay, the model obliges. then the tester says draft an email with this content. BAM! the LLM turns the content from bullets to passages even though it was not asked and it undid the last good thing that it did. i am not sure if it's an email etiquette to not put bullets in the email.
It seems worse in all aspects to the 26B A4B? I would have thought dense models beat MoE still on many benchmarks?
Is the entire point of this model then that it runs if you don’t have enough GPU memory to load the 26B? That one runs faster anyway due to lower active params.
I dunno, feels a bit unfair to companies that actually do FOSS releases (Gemma 4 being released under Apache 2.0 license) to compare them to a company that never done any FOSS releases, and mostly done proprietary "available to download" releases.
Agreed, miles ahead though from "proprietary" which is what Meta been using for most model releases.
Ideally companies would share the fucking datasets and training code already, but no, no one wants to talk about the source of those or even share the ones they have as then who knows what comes out of Pandora's box...
IDK this model release is a bit disappointing considering the community has been chomping at the bit for the 124ba4b model. There was some leaked info about it but people suspect it was not released because it was too close to gemini flash in performance.
Every other Google model I have tried felt very weak compared to qwen models. I dont have a ton of use case for multimodal though, so its very possible this is a fantastic multimodal model.
Hard disagree, Qwen multimodal is way better than google's, but Gemma 31b runs laps around Qwen 27B in complex engineering tasks. Maybe Qwen is better at slopcoding web framework CRUD, but for embedded dev there's no comparison.
Turns out when you block people from the best and biggest hardware, they get innovative. It reminds me of the Pentium days when everyone was shipping inefficient programs because the processor would be better next year.
Since ollama has diverged from llama.cpp, it will take a bit of time for ollama to support multi-modality. If you're using plain llama.cpp it looks like a PR has already merged for this model with vision and audio support:
They've actually gone back to (a lightly patched) llama.cpp with the 0.30 release a few weeks ago, and have now vendored-in an up to date release. Needless to say this is great news for both projects!
Is this Mac only? Or is that an Ollama issue that it only supports this release of models on Mac? It seems like every tag with the MLX badge is only supported on Mac[0], and that includes all of the tags in this release.
MLX is quite literally macOS-specific technology, for other platforms you want non-MLX.
I was sure "MLX" stood for "Metal-something-something" but can't find any reference to that somehow, anywho, "Metal" is hardware-accelerated graphics on Apple platforms FWIW.
Edit: about the actual release on Ollama, if you're on non-Apple hardware you probably want the NVFP4 variant ("gemma4:12b-nvfp4") which was uploaded 45 minutes ago, especially if you're with a recent nvidia GPU.
I realize this is a little confusing; we're working w/ the MLX team to bring MLX to other platforms, but we're not quite there yet. The `gemma4:12b-nvfp4` model is specifically for the MLX engine.
For the GGUF 4bit variant (i.e. non-macs) you'll need `gemma4:12b-it-q4_K_M` which I just pushed. You'll also need to upgrade to version 0.30.4 which we're just about to release (it's in prerelease and we're running through our last regression tests).
Do Gemma 4 models compete with Gemini 3.1 Flash-Lite? I would assume even the smallest Gemini model would outperform even Gemma 4 31B, but I can't really get a sense of performance or output quality difference.
Gemma 4 31b outperformed Gemini 3.1 Flash-Lite in our app benchmarks (agentic tool use via api in our application as a part of various workflows). But google won't let you pay to use Gemma models, you have to go elsewhere, I think this may be because it would cannabilize Flash-lite.
You can actually get the gemma-4 models on a per-token API basis, you just have to click some extra buttons (in GCP). Not the same for other open weight models. For those they make you run your own hardware.
You can get SXM V100s for like $100 off ebay, if you're willing to do the troubleshooting work to get em running with adapters you can build a computer capable of fitting a Q4 quant of a 120b model in VRAM for something like fifteen hundred dollars. (assuming you already have some RAM sticks laying around T___T)
> Novel unified architecture: No multimodal encoders. The vision and audio inputs flow directly into the LLM backbone.
I would be interested in how this actually works. I couldn't find a description of the model architecture (and I did check the links in the Google blog)
My understanding is that early (and most extant) visual language models have a component module (called the image encoder) that transforms images into representations (called embeddings) the model's inner layers can process.
This is often a separate module grafted onto the main model, and further pre-trained (e.g. OpenAI's CLIP, SigLIP used in the Gemma 3 and PaliGemma series).
The image encoder approach has a few problems.
One problem is that many like Gemma 3's encoder have fixed image resolution constraints and inputs must be resized with all the attendant distortions that causes with spatial understanding. However, the Gemma 4 series image encoders overcame this and can handle variable-dimension inputs.
Two, these image encoders are somewhat large (ranging from 300-500M parameters) requiring extra memory and FLOPs to run.
Three, say we need to fine-tune a vision language model, updates to its weights, may affect its understanding of the representations generated by the image encoder if we don't fine-tune both together.
The new Gemma-4-12B replaces the encoder (with its many attention layers and large parameter count) with a simple linear projection to generate the embeddings for images. That reduces the computational requirements and simplifies the input pipelines for image processing.
I don't have any expertise on the topic though and might very well be wrong on some details.
Quite a niche release. The MoE outperforms it on score and will likely be faster thanks to lower active weights. So this really only makes sense for specific ram constrained applications that can’t fit a quantized MoE
I'm seeing very low quality results on LMStudio with this model. Worse than Gemma 3 12B.
It is getting questions like "David has 18 apples and Ivan has 7 apples. How many apples do they have together?" wrong half the time, while Gemma3 12B could very consistently answer that. Other smoke tests (like Chinese translation, and the infamous "Rs in Strawberry" test) also show poor results.
I don't know if it is a quantization/release issue, if the parameters needed for accurate responses have changed (i.e. it needs "thinking" tokens to handle its base error rate), or if the model has been so focused on audio/video that the text processing is bad.
I do enjoy the immediate out of touch signaling with the "runs on your 16gb vram laptop" line. Because everyone has a laptop with 16gb vram, or can just pop out and buy a new one, right?
Consumers were complaining about the standard 8GB with the early 2020 refresh of MacBook Pros, many OSes ago. Sure, it might be workable for many tasks (as evidenced by the recent sales of the MacBook Neo), but users with a mere 8GB shouldn't have expectations of LLM performance. Even 16GB feels like a stretch.
On a Mac they are the same thing; they're shared. Of course you need some amount for the OS, but if you have an Apple Silicon Mac with 24GB of RAM, you can likely run a 16GB model.
Which most people as a matter of fact don't use. A majority of people with laptop have separate memory pools and the VRAM of them is nowhere near that and even on most gaming laptops you aren't getting 16GB VRAM.
The result is decent, but it had a few bizzare/trivial syntax errors I had to fix manually: it would do an extra closing bracket or paren a few times, and wanted to separate function definitions with comma. Not sure what that was about, but otherwise the output run just fine.
So, with those qualifiers, I think it's a decent local coding model. It roughly compares with GPT-4.1 (!!), released 14 months ago, on the output: https://senko.net/vibecode-bench/2025/minesweeper-gpt-4.1.ht... (actually I'd call it better, but those syntax errors...)
I ran the quantized version (4-bit GGUF) on my consumer-grade card with 12G of VRAM and got 5t/s for output. Not for interactive use for coding, but fairly capable model.
To me, it's fascinating how much progress we got in over a year. GPT-4.1 was considered an extremely capable coding model. Now we got something with 12B of params performing roughly the same (in this specific benchmark, disclaimers, etc).
Lists of various models I tested: https://senko.net/vibecode-bench/
For 16GB laptops, Qwen 3.5 9B is the undisputed champ.
Gemma 4 31B is the top dog at small model coding, but is dense so it needs ~48GB unified RAM for full context. If you want decent coding on a laptop you need a lot of RAM. But this shouldn't be surprising, dev machines have always needed lots of resources.
That speed for token output indicates to me that it somehow is using hybrid mode and involving cpu+system ram somehow. That ~5tk/s is about the ram bandwidth of DDR4 RAM versus that size model at 4bit. Any consumer GPU with 12 GB like a nvidia rtx 2080 or rtx 3060 should be doing 20+ tk/s with llama.cpp and CUDA backend.
I think the mayor win for coding was reasoning. That's why such a small model can match GPT-4.1 in coding, but I suspect that GPT-4.1 still wins in general world knowledge due to bigger size.
Encyclopedic knowledge matters relatively little in perspective, given the expectable future developments: even the more knowledgeable of us will use that knowledge for reasoning and intuition (and we will have absorbed the intellectual keys during our training), but under our professional hat we should in theory be ready to go "I stand corrected" and "more precisely" with the actual data at hand.
I.e.: for the encyclopedic knowledge needed, the /understander/ will have a RAG subsystem and a corpus of knowledge to inquire upon processing queries.
(Corroboration: we can't delirate, and neither can the machine...)
Thank you for giving me hope!
> Vision: We replaced Gemma 4’s vision encoder with a lightweight embedding module consisting of a single matrix multiplication, positional embedding and normalizations.
That's technically encoding, just without using a dedicated model for it like SigLIP? The Developer's Guide elaborates, it's still a 35M layer which I am curious is robust enough. https://developers.googleblog.com/gemma-4-12b-the-developer-...
> Small enough to run locally on consumer laptops with 16GB of RAM, it unlocks powerful multimodal and agentic experiences right on your machine.
I am assuming that involves quantization, which due to the quality loss makes that statement somewhat misleading IMO.
FAIR did this 2 years ago now: https://arxiv.org/abs/2405.09818
I've been waiting for something like this to be released since then.
The annoying thing is that chameleon was multi-modal out based on the same principles, but this model is just inputs... (I'm curious how they did pre-training without having multi-modal outputs as well. I wonder if they just chopped them off rather than support image output).
Standard approach for training MM-LLMs is we train the encoder first, there are O(2-10B) good images on the internet, so encoder needs to see each image O(10-100) times, that is O(100T) tokens, which is more than the entire pre-training budget for most runs. That is the reason we train the encoder separately (smaller model, 2B active vs 30B or 200B active LLM); there is nothing magical about training the encoder and LLM together, it is just more token-efficient to train the image modality first.
After that a s1/s2 system: fast generation, slow wave correction / observation operating over the fast generation seems like the next leap forward.
Tokens create and hide too many problems to be the 'optimal' solution.
12b means 12G @ 8 bits/param (basically lossless) and 6G at 4 b/p (generally accepted 'pretty close' level). Not too bad?
But TBD how well the base model performs before thinking too much about quantization
> Audio: We simplified audio processing even further. We removed the audio encoder entirely and projected the raw audio signal into the same dimensional space as text tokens.
From the visual guide, there's still the 35M parameter embedder, then the linear projector, for vision, and the linear projector for audio, so it does have some parameters used for the multimodal input to project it into the LLM latent space: https://newsletter.maartengrootendorst.com/p/a-visual-guide-...
And the Unsloth quants, which are missing this, don't support multimodal input. (edit: actually, I may have just needed to update my llama.cpp, will check with an updated llama.cpp soon)
I'm downloading the ggml-org GGUFs now, I tried Unsloth but got some weird problems, double checking with the bf16 model to see if the issue was just the quant.
https://developers.google.com/edge/gallery
Anyone with a 16GB Mac — that is quite a lot of journalists, surely — can download that, install a model into it, and play.
Surely journalists have to start asking questions at least about OpenAI's consumer revenue projections now.
I am a major, major AI cynic, but I decided to be an informed cynic so I've been playing with local models for agentic work and a bit of CAD-to-image generation. I really quite like the 26B Gemma model — I've been using it to teach myself some fundamental things and learn OpenCode without developing a cloud dependency. It writes fairly good code and it is helping me learn the things I want to learn at a pace that I prefer.
But if this 12B model is even half as close as they say it is, this casts some doubt on the consumer end of the cloud business model, at least in the short term.
(Not clear if this app is using the MTP drafters; I've still not got them working with Gemma myself, though the Qwen 3.6 built-in MTP support is super in LM Studio)
However, on my 18GB RAM MacBook Pro, selecting Gemma-4-12B-it results in this error:
> The model "Gemma-4-12B-it' requires more memory (RAM) than is available on your device.
So yeah, my questions about the 16GB marketing copy are fair.
(Though perhaps it'll squeeze in with a small context window? Not sure I understand that aspect yet)
It does seem to use MTP, yes, and it is quite quick — seemingly the underlying LiteRT stuff can do MTP with Gemma 4 and presumably MTP is a big part of the practicality picture here.
The system prompt thing was a surprise when I poked around.
Still progress, but not quite democratic yet.
Weird though that Google might be cannibalising it's own AI subscription service?
https://github.com/ggml-org/llama.cpp/pull/23398
Please don't use Ollama, it's a bad actor in the OSS community.
But I've moved on from Ollama for the time being, though I am mainly interested to see what the Gemma 4 MTP speeds are like on my M1 Max, so I may test it.
I am quite impressed with the tools in LM Studio, which is also a beautiful app, but it is not open source (which challenges my personal strategy somewhat) and I dread its inevitable enshittification.
Nevertheless the GUI has been very helpful while I learn, and I will probably use it until something else presents or my usage pattern settles down from experimentation to something a bit more routine.
I will try oMLX, too, but judging by the LiteRT page I may soon be able to just use that for the larger models if I end up settling with Gemma 4.
So much to learn but this news has really vindicated my decision to direct my limited span of concentration and focus to learning how to use open weights models and opencode.
https://github.com/baaivision/EVE
Isn't that just projecting the patches into the d_model size vectors that the models takes?
>I am assuming that involves of quantization
12B model in 16GB seems very reasonable to me, int8 is top quality for running models.
12B at int8 would take up 12G memory, or 75% of the system memory which technically fits within 16GB but the OS will not like that. EDIT: On my 18G memory MacBook Pro, LM Studio reports a "partial GPU offload" for the int8 MLX weights. Can't test because the `gemma_unified" architecture is NYI.
The part I hate though is that I’d bet none of the performance claims are based on int8.
Why do we care about bf16 benchmarks when no one will be using that with this model.
It sounds like marketing spin where the performance claims are based on BF16 and the “runs in 16GB” claim is on a totally different quantized version.
I'm both shocked but also not surprised that they continue to develop such efficiencies. Honestly it's like silicon and CPU architecture advancement. We kept shrinking it and shrinking it and it kept getting more and more powerful and here we are with AI and it's only going to be 100x more efficient with time. Maybe there's some point of decay but essentially the next 30 years will be more advanced than the last 30 and were going to be living in some sort of futurist blade runner scenario where gene editing is repairing ageing cells, organs and curing all sorts of cancers that haven't even appeared yet. Beyond our lifetimes people will live to 125 quite steadily and with great mobility and then obviously people will look to how do we get to living 1000 years, which of anyone is religious knows Noah and others lived to that age in a totally different era.
Anyway I'm going off on some tangent but look back 30 years. Now look forward 30 years. It's going to be insane. May God protect us.
It's definitely an exciting time, but in terms of advancements in the state of the art, there is a lot of low-hanging fruit left to pick. There IS a bottom, however, as you can only encode so much "knowledge" in a small number of parameters.
This feels to me a lot like what the early days of what radio or aviation must have been like. Or, heck, microcomputers even.
Is it simply goodwill and/or marketing? Or am I missing something strategic?
If that inference becomes popular and valuable enough that those companies make billions of dollars in profit, those companies could use that profit to fund the building of alternative products and platforms that dis-intermediate google's relationship with the customer.
Google already has an 80% gross margin business, the biggest one in the world. Everybody wants a slice of it.
By offering frontier inference closer to cost and open-sourcing everything that's sub-frontier, they're commoditizing frontier labs' models, which inhibits their ability to durably make high gross margins on inference.
It's a strategic play.
> By offering frontier inference closer to cost *and* open-sourcing everything that's sub-frontier
It's two prongs! One prong is that their frontier inference pricing is significantly cheaper/closer-to-at-cost as Anthropic's.
The subject of this thread is the other prong: offering compelling models that are sub-frontier and self-hostable.
Self-hosting models and at-cost frontier models are the high-end and low-end disruptions, respectively, to Ant/OAI/etc.'s business models.
They need one more than ever now.
This is ridiculously anti-competitive.
2. Every time you search for Claude or ChatGPT, you get presented with an AdWords bidding war.
3. Google is deploying its models in Search, Docs/Drive/Office, YouTube, Chrome, ...
2. I'm not sure what this has to do with the case, unless you're arguing Google has an ads monopoly, in which case the best argument would likely not be that adwords lead to bidding wars because that just sounds like they're selling a product people really want to pay for
3. There's nothing criminal about being a very diversified business
That's my experience right now... my company is all in on a plethora of platform products. Also, Microsoft just yesterday said their goal was "Unmetered intelligence". There's a lot of things that can be enabled by small local models, and those things are part of stacks that can generate revenue in other layers.
Of course it is...
This is Windows-Licensing-Level Money Opportunity 2.0.
So it's easier to just release those models as open source and make it official, since someone would inevitably hack the weights out anyway.
Companies don't commonly give away executable binaries "just because", why'd they start now for these binary blobs that are the models?
Not that I'm unhappy about it! Yay for open data any day, I'm just not understanding why, at least beyond PR in nerd circles
They could lock them down legally which would prevent commercial use, but they choose not to, and they boast about how many tens of millions of times Gemma models have been downloaded by developers.
So there must be more to the rationale than just local model weights getting hacked out of devices.
They rise with the tide of AI adoption. But they gain ground if people opt into Google solutions. And any token sent to a Google model (free or paid) actively punishes their competitors that are then required to spend vast sums to remain bleeding edge.
The question is: do you want to release your models, or use them purely for R&D?
Since everyone else is already releasing models of similar qualities, it's hard to say you're shooting yourself in the foot if you join the chorus.
The added cannibalization of releasing them is effectively zero, so the reputational benefits are likely to be worth it.
Nobody would be looking at Qwen if their ~30b class models weren't fantastically good, it's great advertising and builds significant goodwill with developers, who are going to be your biggest advocates.
The other thing is, all these models are already disposable grade, and in a year they'll all be outclassed by The Next Big Thing. "Open" models are less than 18 months behind SOTA right now and I can't imagine that will slow down much over the next two years, they may even begin to close the gap. Nobody even talks about llama 4 anymore despite only being a year old.
So perhaps another part is just Google showing that they can indeed play at the big boys table.
https://youtu.be/JNyuX1zoOgU?is=PdzCILyi8SP6cfDr
We saw great results in our usecase using google direct. Moved to Openrouter because google wouldn't let us use it beyond a test.
Then Openrouters performance looked worse, not sure if there was a quantized version or something. So we instead looked at Deepseek v4 Flash, and opted to go for that.
This model would probably be great for a super low cost cloud model, would love to use it in the cloud, Google makes you go elsewhere.
I'm pretty sure they are doing it because they get some research experience by shrinking and improving these models, and because they know that by doing this they get some good PR among the dev community.
They remind me a bit of HuggingFace, create something great then make money … maybe.
Plus every open model undermines their local competition by furthering open research and reduces moats, especially since Gemini as a frontier model isn't really competitive with GPT nor Claude for most applications.
Eventually the local model is not enough, and you'll upgrade to the big ones.
It may be that the Q6 quant I got is borked (or my LM Studio is), but either way, the 0.8b's performance is mind boggling in comparison.
A model that comfortably fits in 16GB of VRAM (allowing room for context) is a welcome upgrade.
- Transcribing scanned documents into formatted text
- Captioning/describing images and classifying them for audience suitability (includes anti-spam)
- Matching documents with relevant Wikipedia pages for tagging
I don't use them like frontier models. I break the work down into micro-tasks with one clear goal for each prompt. I write a lot of glue software to make the complete flow work. I was working on all of these tasks before LLMs appeared on the scene. The LLMs have allowed me to replace a lot of complicated code with less code plus a model, while achieving better results.
I use local models for reasons of cost and control. I already had the workstation and GPU. The only running cost is electricity. I have used proprietary models from OpenAI and Google for some of these tasks, but I also encountered churn when the models I built my tools around were retired. I don't worry about that when I have the weights saved locally.
I saw a little app the other day, I think someone posted on here, that looks at your screenshot and renames the file based off the contents of the file.
There's tons of little examples like that. For a lot of use cases, you really don't need the frontier models.
If you have a very specific idea for local model use you can find a way to make it work very well, you don't even need to have a graphics card or NPU chip. You just have to be extremely constrained in how it's used. I think as a generic chatbot they're not great, I'd use a hosted SOTA model and I'm a big fan of local LLMs myself.
Could you talk a bit how you did the finetuning? Did you use unsloth or any other tool and how went the verification to proof the outcome?
Yea absolutely, but man, where to even start, it is very specific.
Fundementally I didn't use any wrappers like unsloth or axolotl, although I have used the latter before a year or two back and it was good, but I needed something very very custom. I also wanted the whole fine tuning pipeline to exported OpenVino model to be seamless.
I heavily leaned on codex, claude and some manual sleuthing around the internet to understand what I needed. I'd played about with QLoRA finetuning with axolotl before and felt most comfortable with that. So I needed to keep everything as stripped down as possible and figured I can just utilise the 3 main huggingface libraries (transformers, peft and datasets) and also bitsandbytes (as suggested by claude to quantize the model to keep this working on my GPU) along with some custom scripts generated by claude/codex (each cross referencing each other) that will do the different stages of the training run.
The next part was the data. Obviously didn't have access to thousands of meetings and associated output documents but I did have a 3090ti sitting there and a codex subscription. So I set about working out what format I needed the data in (many thanks again, to claude/codex) and started generating hundreds of different transcripts, different amounts of speakers, content, tones, subjects, spelling mistakes - like all the different things you could think a meeting would have. Then it's a case of actually generating a good meeting document off the back of the transcripts and creating the "gold standard" that we'd use.
I'm going to gloss over a lot here as I'd rather not detail it as it relates to some propriatary stuff that I had to work through, but you basically pair the transcripts together and run the training.
At the verification stage, there was pretty much 3 things:
1. "just" do some regex string matching to see if there's any of the source transcript key facts in the output to ensure fact preservation. Same with owner fabrication (who said what), I don't want something attributed to someone when it wasn't them that said it and then finally markdown validation.
2. Using codex/claude to validate the transcript and output from the model - I used the latest frontier models, probably overkill for my task, but they were good at the job
3. Finally me going through some actual recordings of myself, groups, meetings and manually verifiying the output
So a fair bit of work, and for context I'm on version 10 now, so it's been a journey!
Repo is https://github.com/Rebreda/listenr - mainly geared toward Whisper fine-tuning, AMD hardware and local inference
Then when I’m getting close to feature-complete, I’ll move to a hosted frontier model for the final integration.
Cost savings are enormous if you’re making dozens of calls to language models a minute.
In practice I haven't got around to building something around multimodality since I'm primarily using their text generation capabilities.
I expect it to be something like https://huggingface.co/OuteAI/Llama-OuteTTS-1.0-1B-GGUF
I'll run some cheat-proof benchmarks ones tomorrow see if qwen is still on top.
Is the entire point of this model then that it runs if you don’t have enough GPU memory to load the 26B? That one runs faster anyway due to lower active params.
Ideally companies would share the fucking datasets and training code already, but no, no one wants to talk about the source of those or even share the ones they have as then who knows what comes out of Pandora's box...
I am not overly impressed with the smaller gemma models. And gemma 3 was a bit of a mixed bag, great at some things, bad at most others
still making my way through deep dives on the chinese open weights, they are all pretty good and way more cost / resource effective
Qwen 3.6 on the other hand barely uses any memory at all for its KV cache.
The underlying LiteRT-LM framework used in the edge gallery does support the MTP drafters for the smaller models, but according to:
https://developers.google.com/edge/litert-lm/models/gemma-4
> Note: LiteRT-LM supports E2B and E4B models today, with support for larger models coming soon.
So even Google aren't shipping MTP support for the 26B and 31B models yet.
https://github.com/ggml-org/llama.cpp/pull/24077
This is a good starting issue with a bunch of linked/related
https://github.com/ggml-org/llama.cpp/issues/22746
[0] https://ollama.com/library/gemma4/tags
Edit: MLX being Mac-only is independent of the model being MLX (and therefore Mac) only. The latter is what I am asking about.
I was sure "MLX" stood for "Metal-something-something" but can't find any reference to that somehow, anywho, "Metal" is hardware-accelerated graphics on Apple platforms FWIW.
Edit: about the actual release on Ollama, if you're on non-Apple hardware you probably want the NVFP4 variant ("gemma4:12b-nvfp4") which was uploaded 45 minutes ago, especially if you're with a recent nvidia GPU.
For the GGUF 4bit variant (i.e. non-macs) you'll need `gemma4:12b-it-q4_K_M` which I just pushed. You'll also need to upgrade to version 0.30.4 which we're just about to release (it's in prerelease and we're running through our last regression tests).
Use OpenCode Go instead: https://opencode.ai/go
Wait, *Excluding Chinese language.
This is ... curious.
P.S. Where is gemma 4 124b?
I would be interested in how this actually works. I couldn't find a description of the model architecture (and I did check the links in the Google blog)
This is often a separate module grafted onto the main model, and further pre-trained (e.g. OpenAI's CLIP, SigLIP used in the Gemma 3 and PaliGemma series).
The image encoder approach has a few problems.
One problem is that many like Gemma 3's encoder have fixed image resolution constraints and inputs must be resized with all the attendant distortions that causes with spatial understanding. However, the Gemma 4 series image encoders overcame this and can handle variable-dimension inputs.
Two, these image encoders are somewhat large (ranging from 300-500M parameters) requiring extra memory and FLOPs to run.
Three, say we need to fine-tune a vision language model, updates to its weights, may affect its understanding of the representations generated by the image encoder if we don't fine-tune both together.
The new Gemma-4-12B replaces the encoder (with its many attention layers and large parameter count) with a simple linear projection to generate the embeddings for images. That reduces the computational requirements and simplifies the input pipelines for image processing.
I don't have any expertise on the topic though and might very well be wrong on some details.
But between same (V)RAM requirement 4 bit 26B-A3B and 8 bit 12B it's unclear which one will win, especially given one is MoE and the other dense.
All the launch benchmarks are at 16 bit.
I'm curious how they pre-trained it... I feel like it must have had audio/image output that they chopped off.
I wonder how hard it would be to add it back on.
It is getting questions like "David has 18 apples and Ivan has 7 apples. How many apples do they have together?" wrong half the time, while Gemma3 12B could very consistently answer that. Other smoke tests (like Chinese translation, and the infamous "Rs in Strawberry" test) also show poor results.
I don't know if it is a quantization/release issue, if the parameters needed for accurate responses have changed (i.e. it needs "thinking" tokens to handle its base error rate), or if the model has been so focused on audio/video that the text processing is bad.
Consumers were complaining about the standard 8GB with the early 2020 refresh of MacBook Pros, many OSes ago. Sure, it might be workable for many tasks (as evidenced by the recent sales of the MacBook Neo), but users with a mere 8GB shouldn't have expectations of LLM performance. Even 16GB feels like a stretch.
Majority of people with laptop have RAM and igpu using some of that as VRAM.