Gemma 4 12B
Google · google/gemma-4-12B-it
Q4 file 7.1 GB (Q4_K_M) · runs well from 12 GB · Weights license: Apache 2.0
Quality
No result in LiveBench 2026-06-25, LMArena 2026-10-22 Not in LiveBench 2026-06-25 or LMArena 2026-10-22
Reported by the publisher
- MMLU Pro
- 77.2%
- AIME 2026 no tools
- 77.5%
- LiveCodeBench v6
- 72.0%
- Codeforces ELO
- 1,659
- GPQA Diamond
- 78.8%
- Tau2 (average over 3)
- 69.0%
- HLE no tools
- 5.2%
- BigBench Extra Hard
- 53.0%
Every reported result
Across Gemma 4 sizes
| Benchmark (publisher settings) | E2B | E4B | 12B | 26B A4B | 31B |
|---|---|---|---|---|---|
| MMLU Pro | 60.0% | 69.4% | 77.2% | 82.6% | 85.2% |
| AIME 2026 no tools | 37.5% | 42.5% | 77.5% | 88.3% | 89.2% |
| LiveCodeBench v6 | 44.0% | 52.0% | 72.0% | 77.1% | 80.0% |
| Codeforces ELO | 633 | 940 | 1,659 | 1,718 | 2,150 |
| GPQA Diamond | 43.4% | 58.6% | 78.8% | 82.3% | 84.3% |
| Tau2 (average over 3) | 24.5% | 42.2% | 69.0% | 68.2% | 76.9% |
| HLE no tools | This size's card does not report this result under these settings | This size's card does not report this result under these settings | 5.2% | 8.7% | 19.5% |
| BigBench Extra Hard | 21.9% | 33.1% | 53.0% | 64.8% | 74.4% |
| MMMLU | 67.4% | 76.6% | 83.4% | 86.3% | 88.4% |
| MMMU Pro | 44.2% | 52.6% | 69.1% | 73.8% | 76.9% |
| OmniDocBench 1.5 (average edit distance, lower is better) | 0.290 | 0.181 | 0.164 | 0.149 | 0.131 |
| MATH-Vision | 52.4% | 59.5% | 79.7% | 82.4% | 85.6% |
| MedXPertQA MM | 23.5% | 28.7% | 48.7% | 58.1% | 61.3% |
| CoVoST | 33.47 | 35.54 | 38.5 | This size's card does not report this result under these settings | This size's card does not report this result under these settings |
| FLEURS (lower is better) | 0.09 | 0.08 | 0.069 | This size's card does not report this result under these settings | This size's card does not report this result under these settings |
| MRCR v2 8 needle 128k (average) | 19.1% | 25.4% | 43.4% | 44.1% | 66.4% |
Values as printed on the card, with the publisher's own settings (card section: Benchmark Results); column “Gemma 4 12B Unified”. Not comparable across publishers and not used in any ranking. Model card
Run it
Run it
File Q4_K_M · 7.1 GB · unsloth/gemma-4-12b-it-GGUF · What is Q4?
Measured 35–152 tok/s · Q4_K_M · 9 machines · Local LLM Advisor · all rows
ollama run hf.co/unsloth/gemma-4-12b-it-GGUF:Q4_K_M- Install Ollama
- Run the command above
- Ask:
Explain what a context window is, in two sentences.
Runs on
Fast memory holds it 90
The planning estimate fits the memory the model can use quickly: graphics card VRAM, or unified memory within the macOS GPU share.
Show all 90
Within memory, outside the fast part 26
Within the machine's memory, outside its fast part: on a graphics card the layers that do not fit run on the CPU from PC RAM (the listed RAM, or 32 GB assumed); on a Mac, above the macOS default GPU share. Not a speed figure.
Show all 26
One memory tier short 3
The same machines with the next memory size up would hold it in fast memory.
Gemma 4 12B is a text model from Google with 12B parameters. It was released on June 3, 2026 and reads up to 262,144 tokens of context. Its Q4 weights plus working room come to about 9 GB before the machine's reserve; the smallest reference machines whose fast memory holds that are Mac mini M4 · 16 GB unified (~13 GB with its reserve) and RTX 3060 · 12 GB VRAM (~11 GB with its reserve). We collected 24 downloadable GGUF and MLX files from 2 publisher sources; the most downloaded GGUF source is unsloth's. It has no result in LiveBench 2026-06-25, so no quality score is shown.
Hardware profiles grouped by the Q4 planning estimate on each machine (dedicated GPUs reserve 2 GB; unified and system memory 25%, 2 to 8 GB): the fast memory holds it, it is within memory but outside the fast part (layers offloaded to PC RAM, or above the macOS GPU share), or the machine is one memory tier short. Each card opens its hardware profile. Not a measured fit or speed.
On the reference machines
| Where Q4 sits | ||||
|---|---|---|---|---|
| Mac mini M2 · 8 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~11 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| Mac mini M4 · 16 GB unifiedFast memory holds itQ5_K_M65,536 tokens | ~13 GB | Fast memory holds it | Q5_K_M | 65,536 tokens |
| Mac mini M4 · 24 GB unifiedFast memory holds itQ8_0131,072 tokens | ~15 GB | Fast memory holds it | Q8_0 | 131,072 tokens |
| Mac mini M4 · 32 GB unifiedFast memory holds itQ8_0262,144 tokens | ~17 GB | Fast memory holds it | Q8_0 | 262,144 tokens |
| Mac mini M5 Pro · 64 GB unifiedFast memory holds itQ8_0262,144 tokens | ~17 GB | Fast memory holds it | Q8_0 | 262,144 tokens |
| Mac Studio M5 Max · 128 GB unifiedFast memory holds itQ8_0262,144 tokens | ~17 GB | Fast memory holds it | Q8_0 | 262,144 tokens |
| Mac Studio · M5 Ultra · 36-core CPU / 80-core GPU · 512 GB unifiedFast memory holds itQ8_0262,144 tokens | ~17 GB | Fast memory holds it | Q8_0 | 262,144 tokens |
| RTX 4060 · 8 GB VRAMIn memory, outside the fast partNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~11 GB | In memory, outside the fast part | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| RTX 3060 · 12 GB VRAMFast memory holds itQ4_K_M65,536 tokens | ~11 GB | Fast memory holds it | Q4_K_M | 65,536 tokens |
| RTX 5080 · 16 GB VRAMFast memory holds itQ6_K65,536 tokens | ~11 GB | Fast memory holds it | Q6_K | 65,536 tokens |
| RTX 4090 · 24 GB VRAMFast memory holds itQ8_0262,144 tokens | ~11 GB | Fast memory holds it | Q8_0 | 262,144 tokens |
| RTX 5090 · 32 GB VRAMFast memory holds itQ8_0262,144 tokens | ~11 GB | Fast memory holds it | Q8_0 | 262,144 tokens |
| DGX H100 · 8 × 80 GBFast memory holds itQ8_0262,144 tokens | ~25 GB | Fast memory holds it | Q8_0 | 262,144 tokens |
The file column uses the site's quant ladder (Q8_0 to Q3_K_M, MLX 8 to 4-bit on a Mac): the largest collected file whose size × 1.2 plus the context cache and the machine's reserve fits the fast memory. Planning estimates, not measured runs.
Compare Gemma 4 12B using its publisher card. No overall score appears in the LiveBench 2026-06-25 snapshot.
The Q4 weights plus working room come to about 9 GB from 11.95B stated parameters, before the machine's reserve (Runs on adds it per machine). It is not a measured minimum; context, batch size, runtime and quantization change actual memory use.
Versions to download
Files you can download for this checkpoint. The section opens on the most downloaded collected GGUF source, unsloth, at Q4_K_M. On a Mac it opens on the most downloaded MLX source. Sizes are the Hub's published bytes. What is Q4?
unsloth
unsloth/gemma-4-12b-it-GGUF · gemma-4-12b-it-Q4_K_M.gguf · 7.1 GB
Run this file
ollama run hf.co/unsloth/gemma-4-12b-it-GGUF:Q4_K_Mllama-server -hf unsloth/gemma-4-12b-it-GGUF:Q4_K_Mlms get https://huggingface.co/unsloth/gemma-4-12b-it-GGUF- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Quality scores on this page belong to the original checkpoint; these files are not scored or ranked. File sizes read 2026-10-23 · download counts read 2026-10-17 · commands: Hugging Face Hub docs: Use Ollama with any GGUF model · llama.cpp server README (-hf <user>/<model>[:quant]) · LM Studio docs: lms get with a Hugging Face URL · mlx-lm README (mlx_lm.generate --model)
Fine-tunes, adapters and other versions
- Quantized versions
- 367
- Fine-tunes
- 245
- Adapters · merges
- 141 · 11
Browse them on the Hugging Face model tree · read 2026-10-23
Uncensored and abliterated versions
Some community uploads remove the model's safety refusals. Quality may drop, the original publisher does not support them, claims about them are the uploader's, and the original license still applies. This site never recommends or ranks them.
66 derivative repositories have “uncensored” or “abliterated” in their name (read 2026-10-23).
Measured speed
Tokens per second by machine
Output tok/s · 9 machines · fastest 5 of 15 · every row in the table · What is a token?
| RTX 5090 | 152.43 ± 0.68 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
|---|---|---|---|---|
| RTX 5090 | 103.67 ± 0.38 | Q8_0 | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf |
| RTX 4090 24 GB | 101.95 ± 0.08 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
| RTX 5080 16 GB | 94.8 ± 0.31 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
| RTX 5070 Ti 16 GB | 89.15 ± 0.31 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
| RTX 3090 24 GB | 86.54 ± 0.16 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
| RTX 5070 12 GB | 70.13 ± 0.02 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
| RTX 4090 24 GB | 64.22 ± 0.09 | Q8_0 | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf |
| RTX 5080 16 GB | 61.81 ± 0.13 | Q8_0 | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf |
| RTX 4070 12 GB | 58.52 ± 0.10 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
| RTX 5070 Ti 16 GB | 58.41 ± 0.18 | Q8_0 | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf |
| RTX 3090 24 GB | 57.24 ± 0.06 | Q8_0 | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf |
| RTX 3060 12 GB | 39.71 ± 0.04 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
| RTX 4060 Ti 16GB | 35.43 ± 0.03 | Q4_K_M | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf |
| RTX 4060 Ti 16GB | 20.58 ± 0.01 | Q8_0 | llama.cpp CUDA | unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf |
Source: Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0 (checked 2026-10-25) · llama.cpp CUDA · llama.cpp commit 3e4dfa8f3 (b11658) · not estimated
How to run
The memory needed above is the Q4 load before a machine's reserve. Formula: parameters × 0.625 bytes, plus 20% working allowance; each machine then adds its system reserve: 25% of unified or system memory, 2 to 8 GB, or 2 GB per GPU on dedicated cards. This is neither measured VRAM nor checkpoint download size.
Q4 planning estimate only. Check checkpoint format, runtime support, KV cache, context, batch size and offload before choosing hardware.
Documented ways to run it
Copied as printed in the linked source (read 2026-10-18); not run by this site.
pip install -U transformers torch acceleratefrom transformers import AutoProcessor, AutoModelForMultimodalLM
MODEL_ID = "google/gemma-4-12B-it"
# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
MODEL_ID,
dtype="auto",
device_map="auto"
)# Prompt
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Write a short joke about saving RAM."},
]
# Process input
inputs = processor.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
return_tensors="pt",
add_generation_prompt=True,
enable_thinking=False
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Generate output
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
# Parse output
processor.parse_response(response, prefix=inputs["input_ids"])Evidence level: Generic estimate only; no execution reproduced by this site.
- Publisher sampling defaults
- temperature 1 · top_p 0.95 · top_k 64generation_config.json · read 2026-10-23source
- Chat template markers
- tool calls · thinkingmarkers found in the template, not a test of the feature · read 2026-10-23source
Before downloading
- Check the exact checkpoint and license linked here.
- Confirm a compatible runtime and precision or quantization for your platform.
- Allow for context, batch size and any extra media components beyond the weights.
Technical details
What the source and the catalog store for this checkpoint. Missing fields are left out, never zero.
- Checkpoint / version
- google/gemma-4-12B-it
- Task and modality
- language and reasoning · text + image + audio + video
- Count and basis
- 12B · manufacturer specificationgoogle/gemma-4-12B-it model card (Hugging Face) · read 2026-10-17source
- Context
- 262,144 tokens published for this modelgoogle/gemma-4-12B-it model card (Hugging Face) · read 2026-10-17source
- Architecture
- Densegoogle/gemma-4-12B-it model card (Hugging Face) · read 2026-10-17source
- Released
- Gemma releases | Google AI for Developers · read 2026-10-09secondary source
- Runtime and precision
- No local execution verified.
- Languages
- 140+ languages Publisher's statement · read 2026-10-18source
- Commercial use
- Allowed Apache License 2.0 text (SPDX) · read 2026-10-18license text
Every stored field
- Revision checked
- 707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
- Architecture
- Gemma4UnifiedForConditionalGeneration
- Library
- transformers
- Hub task tag
- any-to-any
- Declared base models
- google/gemma-4-12B
- Formats
- transformers, safetensors
- Access
- Open download
- Publisher kind
- Official
- Repository created
- 2026-05-23creation is not the model's release
- Repository updated
- 2026-07-20a commit is not a release date
- Hub downloads (30 days) · likes
- 1,720,062 · 1,690this repository only · read 2026-10-25
- Catalog state
- Hub config or weights found; no inventory benchmark link
Sources
Identity and published specifications come from the checkpoint page. Benchmark scores belong to the named release and do not measure speed on your computer.
- Publisher checkpoint and model card
- Weight license and use terms
- Quality benchmark source · LiveBench 2026-06-25 overall
Catalog reviewed: 2026-10-17. Benchmark release: 2026-06-25; snapshot collected: 2026-09-26.






























































