MiniCPM5-2B vs Qwen3-4B-Instruct-2507

MiniCPM5-2B

A tiny dense cousin for local tool calling and long context, with published training data and an active publisher. OpenBMB's own 4-bit builds need about 2 GB, so it runs on a laptop or phone in vLLM, SGLang, llama.cpp, Ollama or Transformers. The main risks are always-on thinking, no safety results and use conditions beyond Apache 2.0.

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.

MiniCPM5-2B

1. Smallest start is ollama run openbmb/minicpm5-2b (1.6 GB, q4_k_m)
2. On vLLM 0.29.0 or later run vllm serve openbmb/MiniCPM5-2B --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser minicpm5, or tool calls come back as XML text
3. On SGLang 0.5.16 or later add --tool-call-parser minicpm5 --reasoning-parser qwen3; SGLang is OpenBMB's recommended backend for tool calls
4. Tool calls are XML, <function name=...><param name=...>, not Hermes JSON
5. Sample with temperature 1.0, top_p 0.95 and min_p 0.0; llama.cpp's default min_p 0.05 can loop, and repetition_penalty 1.05 helps if it does
6. Expect a <think> block before every answer and leave room in max_tokens; enable_thinking false is accepted by the template but OpenBMB doesn't support it for the 2B
7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented
8. Use the main repository's template; the GGUF, MLX and GPTQ copies predate the 29 September fix

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