Gemma 4 12B IT vs Mistral Large 4

Gemma 4 12B IT

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.

Mistral Large 4

A frontier-sized open cousin for coding agents and defensive security, with strict JSON Schema output, parallel tool calls and three reasoning efforts. It needs a multi-GPU server, about 732 GB for the smallest 4-bit build, and vLLM and SGLang run it only from special images. The main risk is a licence that bars Model-as-a-Service businesses without a separate deal.

Gemma 4 12B 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

Mistral Large 4

1. Serve the NVFP4 build on Blackwell with vllm/vllm-openai:mistral-large-4 --model mistralai/Mistral-Large-4-1T-A52B-NVFP4 --tensor-parallel-size 8 --kv-cache-dtype fp8_ds_mla --tool-call-parser mistral --reasoning-parser mistral --enable-auto-tool-choice; released vLLM doesn't load it
2. On H200 use the FP8 build with tensor parallelism of 16 or 8 by 2 pipelines, per vLLM's recipe
3. Pass reasoning_effort low, medium or high (default high); any other value makes the template raise an error, and there is no off switch
4. Read the reasoning from message.reasoning; max_completion_tokens counts reasoning tokens too
5. Use temperature 1.0 and top_p 0.95 at every effort, per the card
6. Ask for JSON with response_format type json_schema and strict true; the template passes the schema to the model
7. The template rejects a tool message straight after a user message and a conversation that doesn't start with a user or system message
8. Add --speculative-config '{"method": "dspark", "num_speculative_tokens": 3}' to use the embedded draft head
9. If you'll offer it to others as a hosted model, read the licence first; Model-as-a-Service use needs Mistral's commercial licence