Some people here say MiMo-V2.6 is bad with tools and are going back to GLM-5.3-Flash. I spent today running MiMo-V2.6-Flash-RL as the backend for an agent harness, on 2× DGX Spark with vLLM, using the tonyd2wild recipe. Most of the "tool problems" I hit turned out to be serving bugs rather than the model. There are three separate issues, all fixable without touching vLLM's core logic. Details below, in case it saves someone a day. 1. With thinking on, replies after the first turn come back "empty" Symptom My agent (Hermes) kept reporting "No response from provider" and retried the same call 4–5 times. The server log was all 200 OK. Only happened in streaming mode with thinking on, and only once the conversation had any earlier assistant message. Single-turn requests were fine. What actually came back delta.reasoning stayed empty. The reasoning arrived as normal content starting with a literal , and the closing never showed up. Any client that strips … sees an unclosed block and throws the whole reply away. Why it happens MiMo's chat template renders every earlier assistant message as {reasoning}{content}, so any multi-turn prompt contains . With thinking on, the stock template ends the prompt at assistant\n and leaves it to the model to emit . vLLM's streaming path runs is_reasoning_end(prompt_token_ids) over the prompt, finds the from history, and decides reasoning is already over before generation starts. Everything then streams as content. Non-streaming works because it only parses the model's output. A controlled test confirms it: a single turn works; adding one prior "Hello!" from the assistant breaks it. Speculative decoding is identical in both cases, so it's not a multi-token-chunk issue. Fix: pre-open in the generation prompt (credit: issue #1 on the recipe repo). At the end of chat_template.jinja: {%- if add_generation_prompt -%} {{- 'assistant\n' -}} {%- if enable_thinking is false -%} {{- '' -}} {%- else -%} {{- '' -}} {%- endif -%} {%- endif -%} Serve it with --chat-template /path/to/fixed.jinja. Verified, streaming with thinking on: single turn after a plain assistant turn after an assistant turn with reasoning after a tool result on a tool-call turn thinking off 6/6 pass: reasoning only in delta.reasoning, no in content, and tool calls parse. 2. The model silently loses its own earlier reasoning in tool loops Xiaomi's docs say that with thinking on, earlier reasoning must be passed back on assistant messages that made tool calls. Two things drop it: Template: the stock template only reads message.reasoning_content. vLLM returns reasoning in a field called reasoning, so clients that echo back vLLM's own format lose it. vLLM: vLLM only reads message.get("reasoning") (vllm/entrypoints/chat_utils.py). A client that sends reasoning_content (my harness does) has it dropped before the template ever sees it. You can confirm it with /tokenize: put a marker string in reasoning_content on a past assistant message, and it's missing from the rendered prompt. Fixes In the template:{%- set reasoning = message.reasoning_content if message.reasoning_content is string else (message.reasoning if message.reasoning is string else '') -%} In chat_utils.py, one extra fallback (I bind-mount the patched file into the container):reasoning = message.get("reasoning") if reasoning is None: reasoning = message.get("reasoning_content") After both, /tokenize shows earlier reasoning in the prompt under either field name. 3. Every reply is capped at 2,048 tokens unless you send max_tokens The checkpoint's generation_config.json has "max_new_tokens": 2048. With --generation-config auto (the recipe uses it for the sampling defaults), vLLM turns that into the default max_tokens. Any client that doesn't send max_tokens gets 2,048 tokens total, thinking included, so thinking-heavy replies get cut off. Fix: override it. Xiaomi's API allows 128K–131,072 output tokens for V2.6. --generation-config auto --override-generation-config '{"repetition_penalty": 1.05, "max_new_tokens": 131072}' Keep the repetition_penalty 1.05 from the recipe. The recipe author documents "tool-call storms" (hundreds of identical tool calls in one turn) under near-greedy sampling without it. Other things worth knowing The "one-line"