Gemma 4 26B A4B IT vs Qwen3.5-2B

Gemma 4 26B A4B IT

A fast mixture-of-experts cousin for general agent work with tools, optional thinking and image input, under Apache 2.0. Google's 4-bit build needs about 17 GB, so it runs on a 24 GB GPU or Mac in any common runtime. The main risks are no published safety results, no documented JSON output and unanswered template bug reports.

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 26B A4B IT

1. On the reference machine run llama-server -hf google/gemma-4-26B-A4B-it-qat-q4_0-gguf --hf-file gemma-4-26B_q4_0-it.gguf --jinja -c 65536, and add the mmproj file for images
2. With Ollama run ollama run gemma4:26b and raise num_ctx or OLLAMA_CONTEXT_LENGTH, since Google's guide says it defaults to 4K below 24 GiB of VRAM 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
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. Validate tool names and arguments before running them, and check any JSON yourself

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