MiMo-V2.6 Distill Qwen 9B vs Qwen3-4B-Instruct-2507

MiMo-V2.6 Distill Qwen 9B

A 9B cousin tuned on MiMo-V2.6's agent work for coding, terminal tasks and visual coding, released as a base for RL research. It needs about 23 GB in BF16, since Xiaomi ships no smaller build. The main risks are no safety results while a seventh of its tuning is offensive security, and nobody at Xiaomi answering reports.

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.

MiMo-V2.6 Distill Qwen 9B

1. Serve it as the card does, sglang serve --model-path XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B --reasoning-parser mimo, and add --tool-call-parser mimo for parsed tool calls
2. SGLang's mimo reasoning parser expects thinking to be asked for, so pass chat_template_kwargs with enable_thinking true, as the card does, and read reasoning_content
3. Calls come back as function and parameter blocks inside tool_call, the same shape vLLM's and SGLang's qwen3_coder parsers read
4. The generation config samples at temperature 0.6, top_k 20 and top_p 0.95
5. Images and video go in through Qwen3.5's processor; the template's audio tokens have no audio encoder behind them
6. The MTP head isn't in the files, so skip MTP speculative decoding even though config.json lists a layer
7. It's an SFT checkpoint meant for further RL, and discussions report runaway generation on llama.cpp, so cap max_tokens

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