Gemma 4 12B
The middle of the family, and for many people the best fit: it reads text, images, audio and video, holds a 256K context, and runs on a 16 GB laptop.
Released June 3, 2026 · Apache 2.0 · figures from Google's model card ·
Three reasons to pick the 12B
The largest model with audio
The 26B A4B and 31B do not take audio. If you need speech and strong reasoning in one model, this is the top of the range.
It fits ordinary hardware
About 6.7 GB of memory at 4-bit, so it runs on a 16 GB laptop with room left for your other apps.
The full 256K context
The same window as the 31B. The edge models stop at 128K.
Gemma 4 12B at a glance
| Full name | Gemma 4 12B Unified |
| Released | June 3, 2026 |
| Parameters | 11.95B |
| Layers | 48 |
| Design | Dense, encoder-free |
| Context window | 256K tokens |
| Inputs | Text, image, audio, video as frames |
| Output | Text |
| Licence | Apache 2.0 |
| Languages | 140+ |
| Training data to | January 2025 |
| Audio limit | 30 seconds a clip |
| Video limit | 60 seconds at one frame a second |
| Image detail | 70 to 1,120 tokens an image |
| Drafter | 0.4B, for faster decoding |
| Recommended sampling | Temperature 1.0, top-p 0.95, top-k 64 |
From Google's model card, the Hugging Face model page and the release history. The whole family is compared on the models page and in technical specs.
Encoder-free, and why it matters
Most multimodal models bolt a separate vision model, and sometimes an audio model, in front of the language model. The 12B does not. Google's model page says all inputs "flow straight into a single decoder-only transformer, reducing multimodal latency".
Images
48 × 48 patches pass through one 35M-parameter projection. The larger models use a 550M-parameter vision encoder for the same job.
Audio
Raw 40 ms chunks of 16 kHz audio are projected straight in. There is no audio encoder at all; the edge models use a 305M one.
What you gain
Fewer parts to load and less memory spent outside the main model. When fine-tuning, there are no separate encoders to manage.
Design details from the Gemma 4 technical report. The rest of the architecture is explained on the features page.
How much memory it needs
The 4-bit build is the one most people should use.
| Build | Memory for the weights | Runs on |
|---|---|---|
| 4-bit · the usual choice | 6.7 GB | A 16 GB laptop, or a GPU with 8 GB or more |
| 8-bit | 13.4 GB | A 16 GB GPU, or a 24 GB Mac |
| Full precision | 26.7 GB | A 32 GB GPU, or a 48 GB Mac |
Context costs memory too
These figures cover the weights only. A long conversation adds more, so on a 16 GB machine keep the context well below the 256K maximum.
Downloads are a little larger
Ollama lists the QAT 4-bit build at 7.2 GB and LM Studio at 7.40 GB, because packaged files include extra components.
A CPU alone is slow
It will run without a GPU, but for comfortable speed use a graphics card or an Apple silicon Mac.
Memory figures are Google's estimates from the model card; the "runs on" column is this site's guidance. Plan with the VRAM guide; Mac users see Mac setup.
Run Gemma 4 12B
ollama run gemma4:12b-it-qat
# Ask about a picture: put its path in the prompt
ollama run gemma4:12b-it-qat "What is in this image? ./photo.png"
A 7.2 GB download. gemma4:12b also works; the -it-qat tag is Google's quantisation-aware build.
lms get google/gemma-4-12b # 7.40 GB lms load google/gemma-4-12b lms server start # local API on port 1234
Or do it in the app: search for "Gemma", download Gemma 4 12B, load it and chat. No terminal needed.
# pip install -U transformers torch accelerate
from transformers import AutoProcessor, AutoModelForMultimodalLM
processor = AutoProcessor.from_pretrained("google/gemma-4-12B-it")
model = AutoModelForMultimodalLM.from_pretrained(
"google/gemma-4-12B-it", device_map="auto"
)
messages = [{"role": "user", "content": [
{"type": "text", "text": "Summarise the key points of this chart."},
{"type": "image", "url": "chart.png"},
]}]
inputs = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True,
return_dict=True, return_tensors="pt", enable_thinking=False,
).to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
Loading classes from Google's model page; the message layout follows Google's guides. This loads full precision, which needs about 27 GB.
./build/bin/llama-server \ -hf google/gemma-4-12B-it-qat-q4_0-gguf \ --ctx-size 32768 \ --n-gpu-layers 99 \ --port 8080
Pulls Google's official 4-bit GGUF straight from Hugging Face and serves it on port 8080.
Step-by-step help: install guide · how to download · runtime selector. Once it is running, see chat, prompting and the API page. If you downloaded before the July 2026 refresh, pull it again.
The best listener in the family
The 12B has the lowest transcription error rate of any Gemma 4 model, and does it without a dedicated audio encoder.
| Model | FLEURS word error rate |
|---|---|
| Gemma 4 12B | 0.063 |
| Gemma 4 E4B | 0.075 |
| Gemma 4 E2B | 0.090 |
| Gemma 3n E4B | 0.085 |
| Gemma 4 31B and 26B A4B | No audio |
Lower is better. From the technical report. Google's model card reports 0.069 on FLEURS and 38.5 on CoVoST speech translation for the 12B, from a separate run.
Keep clips short
Up to 30 seconds each, mono, 16 kHz, MP3 or WAV. Split longer recordings.
Audio goes after the text
Google recommends placing audio after your instruction, and images before it.
from transformers import pipeline
pipe = pipeline(
task="any-to-any",
model="google/gemma-4-12B-it",
device_map="auto",
dtype="auto",
)
messages = [{"role": "user", "content": [
{"type": "text", "text": "Transcribe this audio."},
{"type": "audio", "audio": "meeting_clip.wav"},
]}]
print(pipe(messages, return_full_text=False))
Adapted from Google's audio guide, with the model name changed to the 12B.
More on audio and video input under capabilities.
How the 12B scores
A step below the two large models on knowledge, maths and coding, level with or ahead of the 26B A4B on agent tasks and long context, and far ahead of the previous generation's flagship.
MMLU Pro
LiveCodeBench v6
| Test | 31B | 26B A4B | 12B | E4B | Gemma 3 27B |
|---|---|---|---|---|---|
| MMLU Pro · knowledge | 85.2 | 82.6 | 77.2 | 69.4 | 67.6 |
| AIME 2026 · maths | 89.2 | 88.3 | 77.5 | 42.5 | 20.8 |
| GPQA Diamond · science | 84.3 | 82.3 | 78.8 | 58.6 | 42.4 |
| LiveCodeBench v6 · coding | 80.0 | 77.1 | 72.0 | 52.0 | 29.1 |
| Codeforces Elo | 2150 | 1718 | 1659 | 940 | 110 |
| Tau2 · agent tasks | 76.9 | 68.2 | 69.0 | 42.2 | 16.2 |
| IFBench · instructions | 76.0 | 72.0 | 74.0 | 44.0 | 32.0 |
| MMMU Pro · vision | 76.9 | 73.8 | 69.1 | 52.6 | 49.7 |
| InfographicVQA | 92.0 | 89.3 | 88.4 | 70.0 | 70.6 |
| RULER at 128K · long context | 96.4 | 89.8 | 91.2 | 86.6 | 66.0 |
| LOFT retrieval at 128K | 79.5 | 66.3 | 66.4 | 58.5 | 8.6 |
From Google's model card and technical report, with thinking on and the highest image detail. Green marks where the 12B beats the 26B A4B. Every test is on the benchmarks page; the previous generation is on the Gemma 3 page.
12B, 26B A4B or E4B?
| Gemma 4 E4B | Gemma 4 12B | Gemma 4 26B A4B | |
|---|---|---|---|
| Memory at 4-bit | 4.5 GB | 6.7 GB | 14.4 GB |
| Context | 128K | 256K | 256K |
| Audio input | Yes | Yes | No |
| Phone build | Yes | No | No |
| MMLU Pro | 69.4 | 77.2 | 82.6 |
| LiveCodeBench v6 | 52.0 | 72.0 | 77.1 |
| Tau2 agent tasks | 42.2 | 69.0 | 68.2 |
| Offered on Google's API | No | No | Yes |
Choose the 12B over the 26B A4B
when you need audio, have 16 GB of memory, or work with long documents and agents, where it holds its own.
Choose the 26B A4B over the 12B
when you have a 24 GB GPU and want the stronger answers on knowledge, maths and code, or you want Google's hosted API.
Choose E4B over the 12B
only when the 12B will not fit: on a phone, or a laptop with 8 GB. The quality drop is the largest step in the family.
Recommendations are this site's reading of Google's figures. Pick any two models to compare in the model comparison; rivals are covered in Gemma 4 vs Qwen 3.5.
Every 12B download
Note the capital B in the Hugging Face names.
| Repository or tag | What it is | Use it for |
|---|---|---|
google/gemma-4-12B-it-qat-q4_0-gguf | Official 4-bit GGUF | Running locally with llama.cpp, Ollama or LM Studio |
google/gemma-4-12B-it | Instruction-tuned, full precision | Serving and fine-tuning |
google/gemma-4-12B-it-assistant | The 0.4B drafter | Faster decoding alongside the main model |
google/gemma-4-12B-it-qat-q4_0-unquantized | QAT weights, not yet quantised | Making your own quantised build, such as MLX |
google/gemma-4-12B-it-qat-w4a16-ct | 4-bit, compressed-tensors format | High-throughput serving on vLLM or SGLang |
google/gemma-4-12B | Base model, no instruction tuning | Your own training only |
gemma4:12b-it-qat | Ollama, QAT 4-bit, 7.2 GB | The quickest start |
google/gemma-4-12b | LM Studio, 7.40 GB | A desktop app with no terminal |
Sizes from Ollama and LM Studio, October 2026. What each file type means is on the models page and under QAT builds; everything else is in the download hub. For shared servers see self-hosting.
What the 12B is not
Not the strongest
On hard maths it scores 77.5 against the 31B's 89.2; on coding, 72.0 against 80.0.
Not for phones
There is no mobile build. On a phone, use E2B or E4B - see mobile downloads and Android.
Not on Google's API
Google's documentation lists only the 31B and 26B A4B. Run the 12B yourself or use another host; see pricing.
Not a long-haul agent
Multi-step agent work is the family's weak spot. Keep task chains short and check each step.
More in limitations and safety. For coding use, see Gemma 4 coding.
Does it really run on a 16 GB laptop?
Yes, with the 4-bit build, which needs about 6.7 GB for the weights. Keep the context length moderate; filling the full 256K window needs far more memory.
Should I use the 12B or the 26B A4B?
The 12B if you need audio or have 16 GB of memory. The 26B A4B if you have a 24 GB GPU and want stronger answers on knowledge, maths and code.
What does "encoder-free" mean?
There is no separate vision or audio model in front of the language model. Images go through one small projection and raw audio is projected straight in.
Is it better than Gemma 3 27B?
Yes, on every test Google reports: 77.2 against 67.6 on MMLU Pro and 77.5 against 20.8 on AIME 2026, in a model less than half the size. It also adds audio, doubles the context and uses Apache 2.0.
Does the 12B have a drafter for faster output?
Yes. Google publishes google/gemma-4-12B-it-assistant, a 0.4B Multi-Token Prediction drafter, alongside the main model.
Can I send it video?
Yes, as frames: up to 60 seconds at one frame a second. Audio from the clip is sent separately, in pieces of up to 30 seconds.
Can I fine-tune it?
Yes. The weights are open under Apache 2.0, and with no separate encoders there is only one model to adapt. Re-check its safety behaviour afterwards.
Where can I try it without installing anything?
Google AI Studio offers the 31B and 26B A4B, not the 12B. To try the 12B itself, run it locally or look for a host on the playground page.