English
·
简体中文
·
Website
·
Guide
·
Architecture
·
Benchmarks
·
Discord
A MiMo-native AI coding agent for your terminal.
Fork of DeepSeek-Reasonix, tuned for Xiaomi MiMo models — 1M context, hybrid-attention MoE, prefix-cache optimized.
> [!TIP]
> **Cache stability isn't a feature you turn on; it's an invariant the loop is designed around.** The same three-pillar architecture that delivers 99.82% cache-hit rates on DeepSeek applies to MiMo — every layer is tuned to the byte-stable prefix-cache mechanic.
> [!NOTE]
> **MiMo-Reasonix** is a fork of [DeepSeek-Reasonix](https://github.com/esengine/DeepSeek-Reasonix) (v0.52.0), re-targeted to Xiaomi MiMo models as the default backend. DeepSeek models remain fully supported as alternatives.
## Why MiMo?
MiMo-V2.5 delivers **Pro-level agentic performance at half the cost** of the flagship tier, with a 1M-token context window and native thinking/reasoning support.
| Model | Params | Context | Pricing (CNY / 1M tokens) | Use Case |
|---|---|---|---|---|
| **mimo-v2.5** ★ default | 1T total / 42B active | 1M | ¥0.7 cache read · ¥7 input · ¥14 output | Daily coding, PR review, refactoring |
| mimo-v2.5-pro | 1T total / 42B active | 1M | ¥1.4 cache read · ¥14 input · ¥28 output | Complex architecture, security audits |
| mimo-v2-flash | 309B total / 15B active | 262K | ¥0.07 cache read · ¥0.7 input · ¥2.8 output | Ultra-cheap explore/research subtasks |
| mimo-v2-omni | — | 262K | ¥0.56 cache read · ¥2.8 input · ¥14 output | Multimodal (images, audio, video, GUI) |
> **MiMo-V2-Flash is open source** — 309B total / 15B active MoE, ranks Top 2 globally among open-source models on agent benchmarks. Coding capability surpasses all open-source models and rivals Claude 4.5 Sonnet, at **2.5% of the inference cost** and **2× faster generation speed**.
## Install
Requires Node ≥ 22. Works on macOS · Linux · Windows.
~~~bash
npm install -g mimo-reasonix
mimo-reasonix code my-project # paste a MiMo API key on first run
~~~
Or one-shot:
~~~bash
cd my-project
npx mimo-reasonix code
~~~
**Short alias:** `mrnx` is available as a shorter command:
~~~bash
npm install -g mimo-reasonix # exposes both `mimo-reasonix` and `mrnx`
mrnx code my-project
~~~
| Command | When |
|---|---|
| `mimo-reasonix` / `mimo-reasonix code [dir]` | The coding agent. **Start here.** |
| `mimo-reasonix chat` | Plain chat — no filesystem or shell tools. |
| `mimo-reasonix run "task"` | One-shot, streams to stdout. |
| `mimo-reasonix doctor` | Health check: Node, API key, MCP wiring. |
### Switching between MiMo and DeepSeek
```bash
mimo-reasonix code --model deepseek-v4-flash # use DeepSeek for a session
mimo-reasonix code --model mimo-v2.5-pro # upgrade to MiMo Pro tier
```
Or persistently in `~/.mimo-reasonix/config.json`:
```json
{ "baseUrl": "https://api.deepseek.com", "model": "deepseek-v4-flash" }
```
## What stays the same
MiMo-Reasonix inherits the full architecture from DeepSeek-Reasonix:
### Pillar 1 — Cache-First Loop
Context partitioned into ImmutablePrefix / AppendOnlyLog / VolatileScratch — keeps the byte prefix stable across turns so both MiMo and DeepSeek prefix caches stay at 99%+ hit rates.
### Pillar 2 — Tool-Call Repair
Flatten (deep schema) → Scavenge (missed tools) → Truncation (JSON repair) → Storm (repeat-call detection). MiMo's pure OpenAI-compatible tool calls eliminate the DSML hallucination cleanup needed for DeepSeek.
### Pillar 3 — Cost Control
Flash-first defaults (mimo-v2.5), auto-compaction of tool results >3000 tokens, model self-report escalation (`<<>>`), and transparent per-turn cost badges.
## How it compares
| | MiMo-Reasonix | Claude Code | Cursor | Aider |
|---|---|---|---|---|
| Backend | **MiMo** + DeepSeek | Anthropic | OpenAI / Anthropic | any |
| Default model | **mimo-v2.5** (1M ctx) | Sonnet 4.6 | GPT-5 / Sonnet | configurable |
| License | **MIT** | closed | closed | Apache 2 |
| Prefix-cache optimized | **engineered** | not applicable | not applicable | incidental |
| Cost profile | **¥7/M input** | $3/M input | subscription + use | varies |
| Web dashboard | yes | — | n/a | — |
| Plan mode · MCP · hooks · skills | yes | yes | yes | partial |
## Configuration
One JSON file at `~/.mimo-reasonix/config.json` plus per-project overrides under `/.mimo-reasonix/`.
| Topic | Quick read |
|---|---|
| [MCP servers](https://github.com/z1243161996/MiMo-Reasonix/blob/main/docs/CLI-REFERENCE.md#mcp) | stdio · SSE · Streamable HTTP |
| [Skills](https://github.com/z1243161996/MiMo-Reasonix/blob/main/docs/CLI-REFERENCE.md#skills) | Markdown playbooks, `inline` or `subagent` mode |
| [Memory](https://github.com/z1243161996/MiMo-Reasonix/blob/main/docs/CLI-REFERENCE.md#memory) | User-private knowledge pinned into prefix |
| [Hooks](https://github.com/z1243161996/MiMo-Reasonix/blob/main/docs/CLI-REFERENCE.md#hooks) | PreToolUse · PostToolUse · UserPromptSubmit · Stop |
| [Permissions](https://github.com/z1243161996/MiMo-Reasonix/blob/main/docs/CLI-REFERENCE.md#permissions) | Per-workspace shell allowlist |
| [Web search](https://github.com/z1243161996/MiMo-Reasonix/blob/main/docs/CLI-REFERENCE.md#search) | Mojeek / SearXNG / Metaso |
## Documentation
- [**Architecture**](./docs/ARCHITECTURE.md) — three pillars: cache-first loop, tool-call repair, cost control
- [**CLI Reference**](./docs/CLI-REFERENCE.md) — every subcommand, slash command, keybinding
- [**Benchmarks**](./benchmarks/) — τ-bench-lite harness, transcripts, cost methodology
- [**Website**](https://github.com/z1243161996/MiMo-Reasonix)
- [**Contributing**](./CONTRIBUTING.md) · [**Code of Conduct**](./CODE_OF_CONDUCT.md)
## Community
> MiMo-Reasonix is a community fork of DeepSeek-Reasonix. Original project: [esengine/DeepSeek-Reasonix](https://github.com/esengine/DeepSeek-Reasonix).
Scoped starter tickets under the [`good first issue`](https://github.com/z1243161996/MiMo-Reasonix/labels/good%20first%20issue) label.
## Acknowledgments
Built on [DeepSeek-Reasonix](https://github.com/esengine/DeepSeek-Reasonix) by [@esengine](https://github.com/esengine) and the open-source community. MiMo model information sourced from [LobeHub model-bank](https://github.com/lobehub/lobehub).