Gemma 4 31B IT vs Qwen3.5-2B

Gemma 4 31B IT

A dense multimodal cousin with documented tool calling, optional thinking and a 256K context, for general agent work and image reading on one 24 GB GPU through Google's 4-bit build. It runs in every major runtime under plain Apache 2.0. The main risk is that Google publishes no safety results and has gone quiet on the repository since July.

Qwen3.5-2B

A small vision-language cousin for prototypes, fine-tuning and light local steps such as tool calls, tagging or reading an image, running in about 2 GB at 4 bits in every major runtime under plain Apache 2.0. Keep thinking off, since the card warns it loops. The main risk is neglect, with no safety results or answers from Alibaba since March.

Gemma 4 31B IT

1. On the reference machine run llama-server -hf google/gemma-4-31B-it-qat-q4_0-gguf --hf-file gemma-4-31B_q4_0-it.gguf --jinja -c 32768, adding gemma-4-31B-it-mmproj.gguf for images
2. Or ollama run gemma4:31b for the 20 GB Q4_K_M build
3. In vLLM serve with --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4, as vLLM's recipe does
4. Thinking is off by default, so pass enable_thinking: true in chat_template_kwargs to turn it on, and expect the reasoning in the thought channel block
5. Drop earlier thoughts from the history except on tool-call turns, as the card asks
6. Use the card's sampling, temperature 1.0, top_p 0.95, top_k 64
7. Validate function names and arguments before running a call, and check JSON output yourself or use the runtime's guided decoding
8. The w4a16 build relies on Marlin INT4 kernels that vLLM marks unsupported on AMD, so use the GGUF on the reference machine

Qwen3.5-2B

1. Smallest start is ollama run qwen3.5:2b-q4_K_M (1.9 GB); plain qwen3.5:2b pulls the 2.7 GB Q8_0 build
2. Full precision vllm serve Qwen/Qwen3.5-2B --port 8000 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder
3. Thinking is off by default; if you turn it on with "chat_template_kwargs": {"enable_thinking": true}, stream and cut off runaway output, since the card warns of thinking loops, and add --reasoning-parser qwen3 in vLLM or SGLang to split the reasoning out
4. There's no generation_config.json, so pass the card's sampling yourself, for example temperature=1.0, top_p=1.0, top_k=20, presence_penalty=2.0 for text without thinking
5. Use it for narrow steps or as a fine-tuning base, not as the main agent, as the card itself suggests
6. Validate any JSON it returns, since structured output isn't documented