GuanxuJi / dsh-llm-mimo

dsh-llm-mimo

Xiaomi MiMo v2.5 adapter for DeepSeek Harness LLM seam.

Features

• Thinking chain support: Correctly recognizes MiMo's reasoning_content field and translates it into harness ReasoningBlock format
• Multi-turn conversations: Preserves reasoning_content in assistant messages on tool-call turns, as MiMo requires
• Tool calling: Tool results are serialized to OpenAI-compatible role: 'tool' messages; text-less assistant turns send "" (never null)
• OpenAI-compatible API: Uses MiMo's OpenAI-compatible chat completions endpoint
• Streaming: Full SSE streaming support with proper block assembly

Supported Models

Model	Description
mimo-v2.5-pro	MiMo-V2.5-Pro — flagship model for complex reasoning tasks
mimo-v2.5	MiMo-V2.5 — general-purpose model

Installation

Build

The package follows the harness package layout: pnpm build emits lib/ (JS) plus lib/types/ (declarations), which main/exports/files reference.

pnpm install
pnpm build
pnpm test        # vitest
pnpm typecheck

As a plugin in a DeepSeek Harness profile

A running dsh is composed from ordered layers (the profile's dsh.profile.bundles, then the profile's cordis.patch.yml, then the home-level patch, then any --patch overlay) — a plugin package's own cordis.yml is NOT auto-discovered. To use this adapter:

# From the profile directory (e.g. ~/.dsh/profiles/web), install the package:
pnpm add dsh-llm-mimo

Then append this block to the profile's cordis.patch.yml:

- insert:
    - id: llm-mimo
      name: 'dsh-llm-mimo'
      config:
        apiKeyEnv: MIMO_API_KEY
        baseURL: https://api.xiaomimimo.com/v1
        enableThinking: true

In the monorepo workspace, the package sits under packages/llm/llm-mimo and is referenced by name the same way; the repo's tsconfig base replaces the standalone tsconfig.json shipped here.

Configuration

Set your MiMo API key:

export MIMO_API_KEY="your-api-key-here"

Or in .env file:

MIMO_API_KEY=your-api-key-here

Get your API key from MiMo Platform.

Usage

With DeepSeek Harness Web UI

# From DeepSeek Harness checkout
pnpm dsh web

Then select mimo-official as the provider and choose a model.

Programmatic usage

import { Context } from '@deepseek-ai/cordis'
import llmMimo from 'dsh-llm-mimo'

const ctx = new Context()
ctx.plugin(llmMimo, {
  apiKeyEnv: 'MIMO_API_KEY',
  baseURL: 'https://api.xiaomimimo.com/v1',
  enableThinking: true,
})

Configuration Options

Option	Type	Default	Description
apiKeyEnv	string	MIMO_API_KEY	Credential reference (environment-variable name) for the API key
baseURL	string	https://api.xiaomimimo.com/v1	API endpoint base URL
enableThinking	boolean	true	Whether thinking mode is enabled by default
maxTokens	number	256000	Default per-request output token cap
defaultContextWindow	number	1000000	Default context window size
models	MiMoCatalogModel[]	V2.5, V2.5-Pro	Advisory model catalog
streamIdleTimeoutMs	number	300000	Stream idle timeout (ms)
retryPolicy	RetryPolicyConfig	—	Custom retry policy

How Thinking Chain Works

MiMo v2.5 uses the reasoning_content field in SSE responses to stream thinking chain content. This is the same format used by DeepSeek's thinking mode.

Wire format

{
  "choices": [{
    "delta": {
      "reasoning_content": "Let me think about this...",
      "content": null
    }
  }]
}

Harness translation

The adapter translates this into harness StreamChunk protocol:

// 1. Block start
{ type: 'block-start', index: 0, blockType: 'reasoning' }

// 2. Reasoning deltas
{ type: 'reasoning-delta', index: 0, text: 'Let me think about this...' }

// 3. Block end
{ type: 'block-end', index: 0, block: { type: 'reasoning', text: '...' } }

// 4. Text content (if any)
{ type: 'block-start', index: 1, blockType: 'text' }
{ type: 'text-delta', index: 1, text: 'The answer is...' }
{ type: 'block-end', index: 1, block: { type: 'text', text: '...' } }

Multi-turn conversations

MiMo requires reasoning_content to be preserved in assistant messages when the assistant turn carried tool calls and thinking mode is enabled. On tool-call-free turns the field is ignored by the provider, so this adapter omits it there to save tokens.

API Compatibility

MiMo-specific parameters

• enable_thinking (boolean): Controls thinking mode (non-OpenAI standard, passed via extra_body in Python SDK)
• max_completion_tokens: Used instead of max_tokens for output token limit

Differences from DeepSeek API

Feature	DeepSeek	MiMo
Thinking toggle	thinking.type: 'enabled'	enable_thinking: true
Reasoning effort	reasoning_effort: 'low'	Not supported
Stop sequences	No limit	Max 4
Temperature range	[0, 2)	[0, 1.5]
Top-p range	(0, 1]	[0.01, 1]

Error Handling

The adapter maps MiMo HTTP error codes to harness error codes:

HTTP Status	Harness Code	Description
401, 403	AUTH	Authentication failure
400	INVALID_REQUEST	Bad request (including missing reasoning_content)
429	RATE_LIMIT	Rate limit exceeded
500+	SERVER	Server error

License

MIT