Gemma 4 12B IT vs Qwen3-4B-Instruct-2507

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.

Qwen3-4B-Instruct-2507

A small, fast cousin for instruction following, short tool-calling jobs and long documents on modest hardware, with no thinking step. It runs in about 3 GB at 4 bits on every common runtime, and two providers host it. The main risk is neglect, with no file changes in over a year, no answers from Alibaba and no safety results.

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

Qwen3-4B-Instruct-2507

1. Quickest start is ollama run qwen3:4b-instruct, 2.5 GB. Plain qwen3:4b is the Thinking-2507 edition, not this one
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes
3. Use the hermes parser in vLLM or qwen25 in SGLang, the template emits JSON inside tool_call tags
4. Don't pass enable_thinking, this edition never thinks and the card says the flag isn't needed
5. Sampling per the card is temperature 0.7, top_p 0.8, top_k 20 and min_p 0, with presence_penalty up to 2 against repetition
6. Drop the context to 32,768 tokens if memory runs out, as the card suggests
7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented