Fix max_tokens for Xiaomi

Fix max_tokens for Xiaomi | Generated by AI

Home 2026.06

OK here’s the full picture after inspecting your config and the codebase.

Your Setup

Config: ~/.hermes/config.yaml

Model: mimo-v2.5-pro
Provider: xiaomi
Base URL: https://token-plan-cn.xiaomimimo.com/v1
No max_tokens configured anywhere — not in model:, not in agent:, not in .env

Provider profile: plugins/model-providers/xiaomi/__init__.py

default_max_tokens is not set (defaults to None)

Root Cause

The xiaomi provider profile has no default_max_tokens. Your config also doesn’t set one. So when the agent calls the Xiaomi API, no max_tokens parameter is sent at all — the server picks its own default.

The server’s default output cap for mimo-v2.5-pro appears to be relatively low. When the model generates a response that exceeds that server-side limit, the API returns finish_reason="length", and Hermes reports:

Error: Response truncated due to output length limit

The retry mechanism (up to 3 attempts with boosted max_tokens) also doesn’t help much because the boost starts from agent.max_tokens which is None, falling back to 4096 as the base:

# conversation_loop.py line 1741
_tc_boost_base = agent.max_tokens if agent.max_tokens else 4096
_tc_boost = _tc_boost_base * (truncated_tool_call_retries + 1)

So on retry it sends max_tokens=8192, then 12288, then 16384 — which may still be under the server’s actual limit.

Fix

Add max_tokens to your config. The opencode-zen plugin already knows mimo-v2.5-pro supports 131072 completion tokens. Set it in config.yaml:

model:
  default: mimo-v2.5-pro
  provider: xiaomi
  base_url: https://token-plan-cn.xiaomimimo.com/v1
  max_tokens: 131072    # <-- add this line

Or more conservatively if you want to save tokens:

  max_tokens: 32768

Want me to add it to your config?

Back Donate