MiniCPM-V 4.7 35B-A3B vs Qwen3.5-397B-A17B

MiniCPM-V 4.7 35B-A3B

No score change since it was listed.

Qwen3.5-397B-A17B

• -0.5 on 2 November 2026: time

MiniCPM-V 4.7 35B-A3B

A vision-language cousin for images and video that fits a 24 to 32 GB machine from OpenBMB's own GGUF, under plain Apache 2.0, with a publisher that answers. The card says little about limits, training data or how tool calls come back, so test the tool format in your runtime before relying on it.

Qwen3.5-397B-A17B

Alibaba's biggest open Qwen3.5, a cousin for hard coding, agent and vision work. Even at 4 bits it needs about 270 GB, so most agents will call it through Alibaba's Model Studio or one of the many providers that host it. The main risk is silence, with no safety results, no agent guidance and template reports without a reply.

MiniCPM-V 4.7 35B-A3B

1. Start it with llama-server -hf openbmb/MiniCPM-V-4.7-35B-A3B-gguf:Q4_K_M --jinja, or ollama run openbmb/minicpm-v4.7:35b-a3b
2. For image input in llama.cpp add mmproj-model-f16.gguf, or the 4x projector for more visual detail at a higher cost
3. The GGUF files at the repository root have thinking off. Use the files in thinking/ when you want the trace
4. With Transformers pass enable_thinking to apply_chat_template, and pass downsample_mode to both apply_chat_template and generate
5. Use temperature 0.7, top_p 0.95 and top_k 100 in instruct mode, as the card recommends
6. Check tool calls in your runtime first, since the card names no parser. The template uses the Qwen3.5 function-block format
7. Validate JSON output yourself, since structured output isn't documented

Qwen3.5-397B-A17B

1. The card's server command is vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, and the FP8 build halves the memory
2. It doesn't fit our 128 GB reference machine at any precision, so call a hosted endpoint instead (Model Studio's Qwen3.5-Plus, or a provider serving Qwen/Qwen3.5-397B-A17B)
3. On Model Studio pass "enable_thinking": false directly, while vLLM and SGLang take "chat_template_kwargs": {"enable_thinking": false}
4. Sampling per the card is temperature 0.6, top_p 0.95 and top_k 20 when thinking, and temperature 0.7, top_p 0.8, top_k 20 and presence_penalty 1.5 without
5. For multi-token prediction in vLLM add --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
6. Keep the context at 128K or more if you rely on thinking, and turn on YaRN only past 262,144 tokens
7. Validate any JSON it returns, since structured output isn't documented