MiMo-V2.6-Flash Tool-Call Failures Are vLLM Bugs: Empty Streaming Replies, Lost Reasoning, Hidden 2048-Token Cap
mamolengo · reddit · 2026-09-23
Running MiMo-V2.6-Flash-RL on 2× DGX Spark with vLLM as an agent backend, the author found most reported tool problems are serving bugs, not model flaws:
- Empty streaming replies with thinking on: MiMo's chat template renders past assistant turns as <think>{reasoning}</think>{content}, so vLLM's streaming path finds a </think> in the prompt and decides reasoning is already over — reasoning streams as content and clients that strip <think> blocks discard it. Fix: pre-open <think> in the generation prompt via a patched chat template. Verified across six scenarios.
- Lost reasoning in tool loops: the template only reads message.reasoningcontent while vLLM returns reasoning, and vLLM's chatutils.py only reads reasoning, dropping reasoningcontent before the template sees it. Fix: field fallbacks in both template and chatutils.py (bind-mounted into the container). Confirmable via /tokenize.
- Hidden 2,048-token output cap (details truncated in the post).
Bottom line: audit the serving layer before blaming the model or switching to GLM-5.3-Flash.
More from coding & agent
- Perplexity's hint-guided self-distillation cuts its agent's tool-call failures by 21.2% — perplexity_ai · 2026-09-23
- Tomo ships WebMCP Registry: open index of agent-ready tools across 100K sites — Jackyhuang · 2026-09-23
- LiteParse hits 2.8ms per PDF page, 25% faster in v2.14.6, claims fastest open-source parser — llama_index · 2026-09-23
- UCLA Releases ACLArena: A Framework for Agent Continual Learning in Multi-Stage Post-Training — UCLA-SCAI · 2026-09-23
- The Hidden Cost of Using Weaker Models for Decisions: You Never Benchmark, So You Never Notice the Lost Alpha — generativist · 2026-09-23
- Coinbase opens stock trading to AI agents, with x402 micropayments for live market data — MurrLincoln · 2026-09-23