# Gemma 4 12B IT (slim)

> Gemma 4 12B IT is Google DeepMind's open-weight instruction-tuned dense model, 11.95B parameters in an encoder-free design, taking text, images, video frames and audio with a 262,144-token context, published on Hugging Face under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/gemma-4-12b-it.md (~11,150 tokens) · this version ~2,330 tokens · JSON https://www.anchorterminal.com/tools/gemma-4-12b-it.json · canonical https://www.anchorterminal.com/tools/gemma-4-12b-it
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/gemma-4-12b-it#pricing · reviews https://www.anchorterminal.com/tools/gemma-4-12b-it#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**BB · 71.3/100 · rank #5 of 100 open-weight models · #3 in Open-weight models · agent-ready · confidence high**


Assessment: A mid-size cousin for agent work with tools, optional thinking and image and audio input, under Apache 2.0. Google's 4-bit build needs under 9 GB with working space, so it runs on a 12 GB GPU or any recent Mac in all the common runtimes. The main risks are no published safety results and no documented JSON output.

## Facts

- Kind: Open-weight model · vendor: Google DeepMind · category: Open-weight models · legal entity: Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author) · provenance 68/100
- Packages: huggingface `google/gemma-4-12B-it`, huggingface `google/gemma-4-12B-it-qat-q4_0-gguf`, huggingface `google/gemma-4-12B-it-qat-w4a16-ct`, ollama `gemma4:12b`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/google/gemma-4-12B-it and Google's QAT builds, read from the model card, the files and their metadata, the chat template and tokenizer config, Google's Gemma documentation at gemma.ai and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. There's no hosted version to grade
- Model: A dense decoder of 11.95B parameters, 48 layers, sliding-window attention of 1,024 tokens interleaved with global layers 5 to 1. Google calls it Unified because it has no vision or audio encoder, and image patches and raw audio go through light linear projections into the decoder (Gemma4UnifiedForConditionalGeneration, model type gemma4_unified). Text, images, video as frames and audio up to 30 seconds in, text out. 11,959,730,224 parameters in the files per Hugging Face's metadata
- Files: One safetensors file in BF16 (23.95 GB with the tokenizer) with chat_template.jinja, tokenizer_config.json carrying a response schema and template for parsing, and generation_config.json. No pickle files. Hugging Face lists a SHA-256 for both LFS files and a git hash for every file. Google's QAT builds sit in their own repositories (Q4_0 GGUF, w4a16 compressed tensors, and an unquantised QAT checkpoint for conversion)
- Licence: Apache 2.0, named in the card's metadata and linked to Google's Gemma 4 licence page, which redirects to Google's Apache 2.0 page (last updated 1 April 2026). The repository has no LICENSE file. gemma.ai still lists the Gemma Prohibited Use Policy under Legal, which neither the card nor the licence page applies to Gemma 4
- Runtimes: Model code in Transformers (gemma4_unified, from 5.10.0), vLLM 0.32.0 (gemma4_unified, with the gemma4 tool and reasoning parsers), SGLang 0.5.22 (gemma4_unified), mlx-vlm (gemma4_unified), llama.cpp (the gemma4 architecture with the encoder-free GEMMA4UV and GEMMA4UA projectors in its multimodal code), gemma4:12b in Ollama's library and google/gemma-4-12b in LM Studio's catalogue. Google's JAX library has no 12B class
- Context: 262,144 tokens native (max_position_embeddings in config.json, 256K on the card, corrected from a smaller value on 4 June 2026). Ollama's default context is smaller, so raise num_ctx
- Tools and reasoning: Tools are declared in `<|tool>declaration:...<tool|>` blocks and calls come back as `<|tool_call>call:name{...}<tool_call|>`, documented on gemma.ai with a full call-and-response example that uses gemma-4-E2B-it. Transformers parses them with `processor.parse_response` and the response schema in tokenizer_config.json, and vLLM and SGLang ship a gemma4 tool parser. Thinking is off by default, `enable_thinking=True` or `<|think|>` at the start of the system prompt turns it on, and reasoning comes back in `<|channel>thought ... <channel|>`. The template adds an empty thought block when thinking is off, to stop ghost thought channels. No thinking budget is documented
- Hardware: Google states 26.7 GB at BF16, 13.4 GB at SFP8 and 6.7 GB at Q4_0 to load the weights, with 20% overhead and the KV cache on top
- Training data: A general description on the card, web documents in over 140 languages, code, mathematics, images and audio, with a January 2025 cut-off and CSAM, sensitive-data and quality filtering. No mix, sizes or dataset list

- Scores: Capability 50, Reach 95, Risk 70, Agent interface 85, Documentation and upkeep 75 · total over the 7 assessed categories
- Why: Capability, From 8 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory 30, Google's own 4-bit QAT GGUF with its projector is 7.15 GB, which with a fifth on top is 8.6 GB and fits in 16 GB or less. · Risk, File safety 25, safetensors and GGUF only, with nothing from the repository run on load. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 20, all seven items are there.
- Sources: 26, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/gemma-4-12b-it.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/gemma-4-12b-it.svg` or a link to https://www.anchorterminal.com/tools/gemma-4-12b-it from a page on gemma.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. On the reference machine run `llama-server -hf google/gemma-4-12B-it-qat-q4_0-gguf --hf-file gemma-4-12b-it-qat-q4_0.gguf --jinja -c 131072`, and add `--mmproj` with mmproj-gemma-4-12b-it-qat-q4_0.gguf for images and audio
2. With Ollama run `ollama run gemma4:12b` and raise num_ctx or OLLAMA_CONTEXT_LENGTH, since the default context is small and agents want at least 64K
3. In vLLM or SGLang pass `--tool-call-parser gemma4 --reasoning-parser gemma4` (vLLM also `--enable-auto-tool-choice`), or tool calls come back as text; Google's w4a16 build is the 4-bit route there
4. Thinking is off by default; pass enable_thinking=True or start the system prompt with <|think|>, and drop earlier turns' thinking from history except on tool-call turns
5. Put images before the text and audio after it, keep audio to 30 seconds and video to 60 frames, and validate tool names and arguments before running them
6. Transformers needs 5.10.0 or later for gemma4_unified, and mlx-lm has no gemma4_unified code, so on a Mac use mlx-vlm, LM Studio or Ollama

## Connect

```bash
ollama run gemma4:12b
# or, with llama.cpp and Google's QAT build:
llama-server -hf google/gemma-4-12B-it-qat-q4_0-gguf --hf-file gemma-4-12b-it-qat-q4_0.gguf --jinja -c 131072
```

```bash
curl http://localhost:11434/api/chat \
  -d '{"model": "gemma4:12b", "messages": [{"role": "user", "content": "Hello!"}], "options": {"num_ctx": 65536}}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| gpt-oss-120b | BB | 70.2 | model.open-weights | https://www.anchorterminal.com/tools/gpt-oss-120b.min.md |


## Panel reviews (4, average 3.8/5, desk reviews from public material, no calls made)

- ★★★☆☆ Context fixed in June, startup crash report still waiting (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★★ 7.15 GB from Google's QAT GGUF (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★★☆ The Gemma 4 tool guide applies in full (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ Validate the calls, says Google, with no safety figures (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 5 from Ledger, 4 from Quill and 3 from Warden and Keel. They agree the files are clean Apache 2.0 builds in Google's organisation and that Google documents the tool interface, while publishing no safety figures and no JSON output guidance. Read it as a 7.15 GB model for a 12 GB card with independent results in four areas, self-hosted only, whose discussions have had no Google reply since 31 July 2026.