Gemma 4 QAT models: Optimizing compression for mobile and laptop efficiency

https://blog.google/innovation-and-ai/technology/developers-tools/quantization-aware-training-gemma-4/

406 130 2 months ago

minimaxir2• 2 months ago

It's a bit awkward to release Gemma 4 12B (https://news.ycombinator.com/item?id=48385906), and then a canonical Q4_0 Gemma 4 12B a couple days later.

It's good that this post lists the expected VRAM usage for the models with Q4_0 Gemma 4 12B being 6.7GB, which will indeed fit Google's claims of fitting within 16GB comfortably, altough it confirms that only the quantized version will do so.

Relatedly, in Google's newly released Edge Gallery for macOS, Gemma 4 12B is explicitly listed as unsupported due to not enough RAM even on a 16GB machine, but given the expected VRAM usage here the Q4_0 variant definitely should fit and Google should fix that.

netdur3• 2 months ago

not sure if I understand you, but 4Q and QAT 4Q are different

refulgentis4• 2 months ago

It's super annoying when you have products that utilize these because there's...4? releases in 3 weeks?

- Gemma 4 2B/4B/27BE3B/31B

- Gemma 4 2B/4B/27BE3B/31B x "assistant" / MTP drafter models (i.e. multitoken prediction)

- Gemma 4 12B (2 days ago? 1?)

- Gemma 4 QAT 2B/4B/12B/27BE3B/31B x "assistant" models (i.e. multitoken prediction)

It probably sounds silly and really whiny in the abstract. It just causes a ton of work / confusion downstream that feels unnecessary.

Extremely glad for the output, not glad to have to chase it.

ex. llama.cpp currently supports the originals but not the MTP predictors but there is a patch for the MTP predictors but not for the small MoE models and I think it supports the 12B but maybe not media for it yet and now we have these too and the blog says there's GGUFs (llama.cpp models) but there isn't in any of the 12? repos I clicked through. and ~every consumer-facing local LLM app is built on llama.cpp or a fork of it.

Also if anyone at Google is taking feedback over to b/ or product, pleaseeee stop the "E"2B "E"4B thing, unless it's actually taking up less RAM on Android during CPU inference. I can't tell if I need to treat the 4B like an 8B (i.e. beyond most consumer hardware without a GPU) or a 4B (i.e. will run on most consumer hardware since 2021)

EDIT: And, yes, the QAT 12B x mmproj does not work with llama.cpp. I'm glad there's people who have the luxury of not having to, well, actually use these and treat me as whining :) I'll need to schedule another 4-8 hours of work for the 4th time, no fun!

Aurornis• 2 months ago

I'm not sure why you think it's awkward to have multiple releases. It's better to release models and variations as they're ready, not withhold them all until everything is ready to release all at once.

The Q4_0 is a quantization aware training checkpoint. It's not a simple quantization of the original Gemma 4 12B.

netdur3• 2 months ago

had a good run with Gemma 4 E2B Unsloth 4Q: https://youtube.com/shorts/XLsAnz5aAAI

The E4B model doesn’t fit on my phone TPU, so it swaps to RAM, the QAT version means more accuracy, good!

prism56• 2 months ago

How do you know it swaps to ram vs on the TPU?

Would be interested in testing this on my pixel.

netdur3• 2 months ago

Because TPU has 2GB and weight + context needs more

ComputerGuru3• 2 months ago

How were you getting anything useful out of that? We found the (unquantized!) E2B model to be completely useless at even the simplest real-world classification tasks.

refulgentis4• 2 months ago

@google.com'ers, there are no GGUFs (blog says there is)

minimaxir2• 2 months ago

Isn’t this it? https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf

refulgentis4• 2 months ago

Ah, nice, ty! My excuse is those repos were added to the collection after my comment, but perhaps not :3

satvikpendem12• 2 months ago

Unsloth's collection as well [0], with their results [1]. Looks like they can get very close to 100% accuracy compared to the BF16 model that is unquantized, and Unsloth's quants are better than the original Google's QAT as posted in the article.

Personal I'm using the 2B model for web search and structured JSON output back via Unsloth Studio and its API, works very well for that even with the model embedded on phones.

[0] https://huggingface.co/collections/unsloth/gemma-4-qat

[1] https://unsloth.ai/docs/models/gemma-4/qat#qat-analysis

llmoorator• 2 months ago

you misunderstand what that chart shows - it shows BF16 QAT Q4_0, not BF16 regular.

meaning Google quantized the model to 4 bit and stored the result in BF16 format for compatibility and convenience to downstream packers.

Like storing small 8 bit numbers in full 32 bit integers.

So it's not close to 100% of unquantized BF16.

I'm curious if anybody can explain why Google released 4 bit QAT Q4_0 is not exactly 100% of BF16 QAT Q4_0? seems like it should be just bit twiddling, no further quantization to convert between these two packings. Unsloth talks about "lattice alignment" being an issue.

That being said I hate it that smol model makers, like Google, Qwen, ... only show the BF16 benchmarks when they release a new models, knowing that what people really run are 4-8 bit quantizations, so it's really hard to understand how much you lose when you run 4 bit vs 6 bit...

satvikpendem12• 2 months ago

Ah I see, thanks for the clarification.

coder5432• 2 months ago

meaning Google quantized the model to 4 bit and stored the result in BF16 format for compatibility and convenience to downstream packers.

You also misunderstand what is happening. Google did not do that. Google further trained the original model with an objective of minimizing error when quantized to 4-bit. The BF16 QAT is not an upscaled 4-bit model. When quantized to 4-bit, it should lose less accuracy than a typical 16-bit model loses when quantized to 4-bit, but the loss is not zero, because it is not based on a 4-bit model.

The Gemma 3 QAT report was a bit clearer:

https://developers.googleblog.com/en/gemma-3-quantized-aware...

"Instead of just quantizing the model after it's fully trained, QAT incorporates the quantization process during training. QAT simulates low-precision operations during training to allow quantization with less degradation afterwards for smaller, faster models while maintaining accuracy. Diving deeper, we applied QAT on ~5,000 steps using probabilities from the non-quantized checkpoint as targets. We reduce the perplexity drop by 54% (using llama.cpp perplexity evaluation) when quantizing down to Q4_0."

The BF16 is just trained to be more resistant to simulated quantization, which helps when it is actually quantized. Google is not doing post-training on the 4-bit model directly.

slopinthebag• 2 months ago

I'm confused, the unsloth model is ~600mb and the one from google is 7gb?

overfeed2• 2 months ago

One is quantized, the other one is Quantization-ready.

scosman• 2 months ago

Google's QAT claims to need 6.7 GB RAM, vs Unsloth's dynamic quants at 8GB. Would love to see some benchmarks. Both amazing for size.

mft_• 2 months ago

Is this [0] saying that unsloth's versions of Google's QAT models are better than Google's own QAT models? Or am I not understanding it correctly?

[0] https://unsloth.ai/docs/models/gemma-4/qat#qat-analysis

ComputerGuru3• 2 months ago

It's saying it's better than naively truncating the QAT release to 4 bits.

cr3cr3• 2 months ago

For a moment I got excited thinking QAT is Intel Quick Assist Technology...

razighter777• 2 months ago

Same I had to do a double take. Would be pretty humourous if they somehow took advantage of crypto offloading to accelerate ai inference

somewhatrandom9• 2 months ago

Could these quantized models make MTP (Multi-Token Prediction) significantly faster when used as drafters for larger regular Gemma 4 models?

dist-epoch2• 2 months ago

Google already released specialized drafters for Gemma 4.

Havoc2• 2 months ago

The E2B ones? Or what do you mean by specialized drafters?

simonw2• 2 months ago

I just ran one of these locally on a Mac like this:

  uvx litert-lm run \
    --from-huggingface-repo=litert-community/gemma-4-E2B-it-litert-lm \
  gemma-4-E2B-it.litertlm \
    --backend=gpu \
    --prompt="Generate an SVG of a pelican riding a bicycle"

The first time you run that it downloads 3.2GB to ~/.cache/huggingface/hub/models--litert-community--gemma-4-E2B-it-litert-lm

It can handle audio and image input too, which is pretty cool for a 3.2GB model. For images:

  uvx litert-lm run \
    --from-huggingface-repo=litert-community/gemma-4-E2B-it-litert-lm \
  gemma-4-E2B-it.litertlm \
    --backend=gpu --vision-backend gpu \
    --attachment image.jpg --prompt describe

And for audio:

  uvx litert-lm run \
    --from-huggingface-repo=litert-community/gemma-4-E2B-it-litert-lm \
  gemma-4-E2B-it.litertlm \
    --backend=gpu --audio-backend cpu \
    --attachment audio.wav --prompt transcribe

(The pelican is rubbish, but it's only a 3.2GB file so the fact it even outputs valid SVG is impressive to me: https://gist.github.com/simonw/94b318afde4b1ce5ff67d4b5d0362... )

reactordev5• 2 months ago

Not to mention the text-only 0.8GB version. Just crazy. You can have basic real-time conversations on-device that's video and audio aware now.

simonw2• 2 months ago

Have you seen a 0.8GB model file floating around yet? I couldn't find one earlier.

yalok• 2 months ago

0.8GB is for text only. It's more like ~1.1GB if you include video/audio encoder

madduci• 2 months ago

Where is it? On ollama I see only the bigger one

viccis• 2 months ago

I'll be honest with you. My main ask for on device AI is that when I am typing "Going out for a quick j" it corrects to "jog" and not "Jonathan". I don't think it needs that many gigabytes.

rcarmo• 2 months ago

Is that actually QAT? the MLX Community models have that in their names, but these don't, and the upload dates don't quite line up.

__mharrison__• 2 months ago

As an aside uvx is so pleasant to use... I wish Nvidia supported it as first-class rather than making folks jump through Docker hoops.

NamlchakKhandro• 2 months ago

I wish people would stop using python sure ai.

It's slow and the PKG resolution is way too flat.

WhiteDawn• 2 months ago

Once someone generates a MTP layer for 26B A4B 4 QAT I'll be singing from the hills with my 5 year old GPU.

dist-epoch2• 2 months ago

Google already did

https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-un...

dofm7• 2 months ago

This is safetensors. Is there any way to run these on a Mac paired with the MLX QAT?

(Pardon my ignorance; this stuff moves so fast)

pfheatwole• 2 months ago

Models:

- Safetensors: https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-un...

- GGUF: https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF/tree/...

Note the README in the Unsloth list of files: llama.cpp is working on a PR to support the gemma4 drafters: https://github.com/ggml-org/llama.cpp/pull/23398. Also note the PR submitter didn't experience much speedup with 26B (seems typical that MoE models don't generally benefit from MTP).

redox996• 2 months ago

I was just testing Gemma E2B and E4B yesterday, and they are just too dumb to be useful outside of niche use cases.

Besides, there's no good agent on Android. Having a model that can't run web searches and browse websites is limited in use, particularly small models that really need to be grounded on search results to be factual, because they can't memorize enough.

Edit: I'd like to know what kind of usage the people that seem to disagree and downvoted this are having.

ilaksh• 2 months ago

I think that's probably true for the vast majority of Android phones. But if you have a SOTA expensive beast, I wonder if Gemma 4 12B at 4 bit could work? Maybe something like a Redmagic 11 pro or OnePlus 13 running NanoClaw?

But also maybe a few Qwen 3.6 or Qwen 3.5 variants can fit and can handle some simple tasks.

redox996• 2 months ago

I think Gemma 4 12B is definitely possible to run on high end phones, google claims you need 16GB of memory. But it's probably not very usable, you'll need to swap most stuff other than the LLM.

When I tried E2B and E4B with Google Edge Gallery, and added a web search skill from the skill list, E2B would fail (get stuck in a loop), E4B would need a very specific instruction, "weather in [city name]" would not call the web search tool, I'd need "web search weather in [city name]". And the result was completely hallucinated and impossible. It claimed 14c and feels like 4c (which is impossible), and 10% humidity (which is almost impossible in this city)

Asking wikipedia level history questions (without any tool use), the results were awful as well.

zkmon• 2 months ago

How can the smaller Unsloth GGUF quant can beat the original google quant? (ref: unsloth/gemma-4-31B-it-qat-GGUF)

SubiculumCode• 2 months ago

I may be wrong, but this is what I figured out. Google provided these quantize-ready models, but they do not come pre-quantized. However, to produce their benchmarks, they quantized their model using the standard quantization approach. Unsloth has an advanced quantization method that performs better than the standard quantization, so the evals are better for unsloth quants.

Catloafdev• 2 months ago

Being able to run the 12B on 8gb VRAM is huge. It's crazy to see how fast these small local models have evolved.

steno1327• 2 months ago

I don't get this obsession with smaller models. I've been using Claude and GPT models for years and have had zero issues with them.

I see absolutely no benefit to me as a end user for a local model which is going to take up more of my CPU and memory and slow down my machine. I almost always have Internet and if I don't then not having access to a AI model is the least of my concerns.

adam_arthur2• 2 months ago

The entire universe of automation projects that can be run effectively for free relative to SoTA models?

I don't think many realize that most LLM embedded automation, pipelines, products will soon be able to run extremely cheaply on models < 100B parameters.

Frontier models will be used for coding/creation use cases, yes. But for all the pseudo-deterministic, pipeline, analysis style things there will be no practical benefit to running frontier models, only additional cost.

Gemma 4 26B outperforms most 100-200B models that I've tested for reasoning and structured output.

Gemma 4 12B can consistently select where to click on browser images given a minimal prompt, and do so very quickly.

steno1327• 2 months ago

Practically if you're running a small personal automation project you're not going to want to waste a lot of time configuring and tuning a local model. You want to build the automation and move on.

If you're building a automation as a company you definitely won't want to take on the long term maintenance overhead of running your own models for some automation project.

dofm7• 2 months ago

The 26B model is really surprising, and it is impressively concise — it spends a lot less time dithering than Qwen3.6.

Zambyte3• 2 months ago

I like using my computer.

steno1327• 2 months ago

Exactly, thank you, we are on the same page! It's great to be able to use our own devices and not have their compute coopted by a third party.

I'd rather not have intensive compute needed shifted onto my personal machine which I want to use for something else.

user2722• 2 months ago

There is tinfoil.sh as well but honestly running this stuff on an airgapped server allows a better peace of mind about the data being used for something else.

steno1327• 2 months ago

What's wrong with the data being used for something else? Someone is providing digital intelligence to us, saving us many hours a week, so the least we can do is provide them a little data so they are able to improve their service.

It would be selfish and unethical not to in my view. And ultimately the data is just being used in order to improve the models and benefit us, not for anything nefarious.

mikeocool• 2 months ago

I've been using Claude and GPT models for years

All 3 years?

steno1327• 2 months ago

GPT1 was released in 2018, so yes, since then.

mannanj6• 2 months ago

I don't like the gaslighting of paying Anthropic or Open(Closed)AI and it being said its unsustainable for them to take my payment while simultaneously they take my data (edit: which is incredibly valuable) and I cannot opt out of that.

The obsession is for leaving hostile and abusive entities, the corporations or the people who fund them that have a horrible track record in regards to ethicality, rights and respect & human dignity.

steno1327• 2 months ago

My view is, if you're going to use the service - you should give the data.

It's like using Gmail and expecting them not to train their AI models on your data - how can you expect that when they're giving you a secure, reliable, highly functional email client completely for free?

The digital economy only works if everyone pays their fair share. If you don't want to give your data then you are really harming everyone by slowing down AI development for everyone else.

sowbug• 2 months ago

Whatever you're doing, try doing 500 or 1,000 of it in a batch. You'll exhaust any subscription quota you have, or if you're paying per token, you will probably find it too expensive. That's when you'll start to ask "how smart a model do I really need for this job?", and you'll investigate running a small but sufficiently capable model on your own PC, churning overnight through your 1,000 tasks.

jbarrow• 2 months ago

Very impressed with how much the Gemma ecosystem has advanced just this week.

Gemma 12B, multitoken prediction, and official quants released. Feels like Google is putting real effort into this string of releases, and I'm very excited to see that!

jhatax• 2 months ago

It’s the Friday before WWDC during which Apple is going to announce an “improved” Siri based on Google models (a locked partnership, for now). Maybe it’s a coincidence, but this might be Google releasing models that will be showcased next week by Apple?

No knowledge, just speculation.

trollbridge• 2 months ago

Maybe Siri will become capable of doing what I can do on my Mac with llamafile and a few minutes of work...

illusive4080• 2 months ago

As an amateur app dev using on device AI: If they replace Apple Foundation model with Gemma 4 I would be so happy.

itake• 2 months ago

I’m curious like what performance if we met you would expect and why?

robgough• 2 months ago

I'm fully expecting an updated foundation model of some kind, but I would bet money they don't utter the words "Google", "Gemini", or "Gemma" even once.

jack_pp• 2 months ago

Ran hf.co/google/gemma-4-12B-it-qat-q4_0-gguf:Q4_0 with ollama on a AMD Ryzen 9 8940HX, NVIDIA GeForce RTX 5060 (8 GB), 14 GB RAM laptop and it is suprisingly fast

Kylejeong21• 2 months ago

google pixel intelligence may beat apple intelligence

nazgul17• 2 months ago

I don't see these QAT models on Edge Gallery; just the BF16 models are there. Is there anything I am missing?

superkuh2• 2 months ago

I wish they would release the base (non instruction tuned) models for use with pattern completion.

llminthefor• 2 months ago

they did

https://huggingface.co/collections/google/gemma-4

superkuh2• 1 month ago

Ah, nice, thanks! https://huggingface.co/google/gemma-4-12B/tree/main No .gguf but I guess I can convert the safetensors myself like it's 2023 again.

RandyOrion2• 2 months ago

From the perspective of a local llm user, I think the qat doesn't solve the major problem of the gemma models.

Gemma family (gen 1 to gen 4) is consistent with extreme range of activations, i.e., 600000, essentially forcing people to use bf16 kv cache and accept a short context window, e.g., 31b, iq4_xs quantization, 100k context window on 32gb memory. Or, people use q8 kv cache, 200k context window, and accept a large performance penalty.

In contrast, for qwen 3.5 family, the largest activation is below 2000, making q8 or even lower-precision kv cache essentially free estates. Together with linear attention, which doesn't require kv cache, full 262k context window can be easily reached.

Qat training with w4a16 target, while improving performance on inference with low-precision weighs, doesn't solve kv cache problem at all.

In the end, a qat is a qat, and there are unseen efforts behind qat checkpoints. Thank you gemma team for releasing qat checkpoints.

RandyOrion2• 2 months ago

More rants about local inference, consider yourself warned.

Together with bf16 related deliberate hardward degrades on consumer-level nvidia gpus, i.e., gtx 10, rtx 20, 30, 40, 50 series, things gets sour really quickly.

nicman23• 2 months ago

the new 4 12b model replaced qwen3.6 27b for me. the task i am doing is a bit specific, validating if a stamp has the correct name but the ones that it could not see maybe a 30 percent were easily discerned.

taffydavid2• 2 months ago

Noob q: can advancements like this targeted at local inference have bonus effects for cloud inference? Presumably if you can get great results on cheaper hardware that also equates to less resource usage on cutting edge hardware, and less power draw?

Will advancements like this ultimately reduce the carbon footprint of AI?

goldenarm• 2 months ago

Consumer and server hardware are quite different, especially Google's TPUs. They notably have much larger mixture-of-experts ratios and more complex caching systems. At such scale and inference budgets, they are incentivised to optimize as much as possible.

Also Google Deepmins has a six month embargo on strategic papers, so I bet the juiciest quantization tech isn't public yet.

arjun-mavonic• 2 months ago

Yet to try this. But from what I heard from a friend is that Gemma 4 12b calls same tool’s repeatedly. Maybe harness can be made to handle it.

make3• 2 months ago

I wonder if they're trying to just kill Anthropic and OpenAI with this, ahah. I assume that Google really doesn't care if the coding model becomes fully overtaken by free models.

linuxhansl• 2 months ago

This is really awesome. I tried my test-task (generate a Python wrapper for a fairly complex C-interface) with gemma4:26b-a4b-it-qat. And for the first time it would just do it, without prodding and without errors. (Of course Claude just did it a heartbeat, too.)

My optimal local setup now is gemmma4-qat and Q8_0 K/V cache quantization with 256k context windows. And that runs fine with 12GB VRAM and another 10GB in RAM.

Previously I tried with gemma4:26b-a4b-it-q4_K_M and qwen3.6:35b-a3b-q4_K_M, and they both would tie themselves into knots (especially qwen3.6 can take forever with incessant "but wait..." thinking loops.) More often than not, they would not finish the task.

It seems true these 4b QAT models are as precise as Q8_0 quantization (which is supposedly indistinguishable from bf16).

I am really excited about the prospect of local LLM inference.

dofm7• 1 month ago

llama.cpp has been patched to support Gemma 4 MTP, so thanks to:

  https://huggingface.co/RachidAR/gemma-4-26B-A4B-it-qat-assistant-q4_0-gguf

you can now do this:

  ./llama-server \
      -hf google/gemma-4-26B-A4B-it-qat-q4_0-gguf \
      --spec-draft-hf RachidAR/gemma-4-26B-A4B-it-qat-assistant-q4_0-gguf:Q4_0 \
      --spec-type draft-mtp \
      --spec-draft-n-max 3

Really quick on an M1 Max, definitely seeing a speedup. Though interestingly the oMLX performance with the MLX community MTP models is much worse. I am not really sure why (the parallel model overhead not being worth the speculative gain, I suppose, but I do not understand this stuff anywhere near as much as I would like).

Gemma 4 QAT models: Optimizing compression for mobile and laptop efficiency

Related Stories