MiniMax-M2.5 vs Qwen3.5-397B-A17B

MiniMax-M2.5

MiniMax-M2.5 is a 230B mixture-of-experts model for coding and agent work, with parallel tool calls, a 196K context and parsers in vLLM and SGLang. It needs a 4-GPU server for FP8 or about 155 GB for a 4-bit build. The main risk is a licence replaced after release with a Prohibited Uses Policy while the card still says modified MIT.

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.

MiniMax-M2.5

1. Start it with vLLM on 4 GPUs of 96 GB, SAFETENSORS_FAST_GPU=1 vllm serve MiniMaxAI/MiniMax-M2.5 --trust-remote-code --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser minimax_m2 --reasoning-parser minimax_m2_append_think
2. In SGLang use --tool-call-parser minimax-m2 and --reasoning-parser minimax-append-think, with --trust-remote-code
3. MiniMax's guides budget 220 GB for the weights and 240 GB per 1M tokens of KV cache
4. On a single large-memory machine, LM Studio's catalogue entry minimax/minimax-m2.5 lists a Q4_K_M GGUF of 138 GB and an MLX 4-bit of 129 GB (Mac only), and LM Studio states 121 GB minimum memory
5. Sample with temperature 1.0, top_p 0.95 and top_k 40, as the card recommends
6. Budget output tokens for thinking on every turn, since the template always opens a think block
7. Read LICENSE-MODEL, not the card's licence link, before shipping, and keep the required Notice text in every copy
8. Without the hardware, call it through Novita or Featherless AI; MiniMax's own API retires M2.5 on 18 November 2026

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