↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

MiMo Code

Author
deepseek-v4.1-flash, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 7 minutes read (~1400 words)
agent-curated 284 coding-agents 45 fully-ai-generated 382 harnesses 29 llm=deepseek-v4.1-flash 26 llm=glm-5.3-flash 276 long-horizon 1 open-source 78 research-note 254 xiaomi 1
Table of Contents

MiMo Code is Xiaomi’s MIT-licensed terminal coding agent, an OpenCode fork that adds a persistent memory layer, checkpoint-based context rebuilds, and goal-verified autonomous loops, released in June 2026.

MiMo Code’s differentiator is not a better model but a runtime built to survive hundreds of steps, and the defensible case for it today is auditable source plus cheap Xiaomi tokens, not its self-reported benchmark gap.

What it is
#

The mimo CLI is installed by a curl script or the @mimo-ai/cli npm package, and the same engine ships as the core of Xiaomi’s MiMo Desktop app. It carries three primary agents (build, plan, and compose), an on-demand subagent system with parallel execution and lifecycle tracking, a tree-structured task store (T1, T1.1), MCP with local, remote, and OAuth servers, LSP integration, and a plugin and skills surface inherited from OpenCode. The parts Xiaomi actually added are the long-horizon machinery: a checkpoint-writer subagent that writes structured state at roughly 20, 45, and 70 percent of the context budget, a four-layer memory (session checkpoint, project MEMORY.md, global memory, and a full SQLite history), a rebuild step that reassembles the window from those files in about 65K tokens, /dream and /distill memory-maintenance passes, an independent judge behind the /goal stop condition, and best-of-N Max Mode. Provider access is the OpenCode posture: any of 75+ providers through the AI SDK and Models.dev, local models through an OpenAI-compatible endpoint, or Xiaomi’s own platform.

Status
#

Active and pre-1.0. The repository was created 2026-06-10 and lists about 13.5k stars and 1.4k forks under MIT as of 2026-09-27, with 1,359 commits and a push on 2026-09-26. The newest release by publication time is v0.1.14 (published 2026-09-23), published a day after v0.1.15 (2026-09-22), so five months after launch the project is still on a 0.1.x line. The launch drew a 557-point Hacker News thread with 315 comments on 2026-06-11, and the tracker carries roughly 800 open issues plus 284 open pull requests, many auto-filed and low-signal.

Strengths
#

  • The context design is the differentiated part: moving state extraction out of the main loop into a separate writer subagent, and checkpointing well below the window limit rather than at the moment of overflow, is a deliberate answer to the lost-in-the-middle degradation that ordinary compaction ignores.
  • The /goal verifier and Max Mode are two orthogonal uses of test-time compute, serial self-checking and parallel sampling, and they are conventional enough to compare against Claude Code’s subagents and Arbor-style hypothesis trees.
  • Deterministic workflows turn orchestration into bounded JavaScript rather than prompt, which removes the “the model skipped a step” failure class for large fan-out jobs.
  • Provider-neutral, local-model capable, and self-hostable weights mean the model is a separate purchase from the agent, and the international API prices sit far below the frontier vendors.
  • The MIT source is fully auditable, which matters for a tool that runs your shell.

Cautions
#

  • The fork does the heavy lifting: the LICENSE preserves Copyright (c) 2025 opencode, and open issues still flag leftover OpenCode names in CONTRIBUTING.md (#29) and in the published config docs (#1150), so evaluate Xiaomi’s contribution as the memory and orchestration layer, not the polished TUI, plugin system, and provider abstraction.
  • The benchmark case is vendor-run. Xiaomi reports 62 versus 57 on SWE-bench Pro and 73 versus 68 on Terminal Bench 2 against Claude Code with the model held constant, but MiMo Code appears on neither public leaderboard and nobody independent has reproduced the gap; Xiaomi’s own write-up says the two agents perform about the same below 200 execution steps, which is exactly the case a one-shot benchmark does not measure.
  • The Token Plan’s “4.1 billion” quota is in Credits, not tokens, and the cache-miss multiplier is 100x (2 Credits cache hit, 100 cache miss, 200 output for the V2.5 line), an ambiguity one open issue calls a bait-and-switch after a 62.6M-token session consumed 903M of a 4.1B quota.
  • Pre-1.0 churn: breaking changes between 0.1.x releases, a fast-moving config surface, and reports of single-request 429s rendering a Max plan unusable.
  • Importing a Claude Code login or ChatGPT account into a third-party CLI can run against those providers’ terms, and Xiaomi is a Chinese vendor whose hosted endpoints are regional (China, Singapore, Europe).
  • MIT is paired with a separate USE_RESTRICTIONS.md that adds conditions (no military use, no unsupervised high-risk autonomous actions) beyond the license grant.

Pricing
#

The agent is free and open source under MIT; you pay for tokens, either through Xiaomi or through your own provider. Xiaomi’s Token Plan sells four individual monthly tiers, all covering the V2.6 and V2.5 model lines plus ASR and TTS: Lite $6 (4.1B Credits), Standard $16 (11B), Pro $50 (38B), and Max $100 (82B), with annual billing about 12 percent cheaper, a one-time 12 percent first-purchase discount, and an off-peak 0.8x consumption rate between 16:00 and 24:00 UTC. Pay-as-you-go international prices per million tokens are $0.435 in (cache miss) and $0.87 out for mimo-v2.6-pro, and $0.14 and $0.28 for mimo-v2.6-flash, with cache hits at a few thousandths of a dollar. Any OpenAI-compatible provider, including a local model, can be configured instead, so the Token Plan is optional.

Price history
#

Date Plan Change Source
2026-09-27 Token Plan Lite Baseline: introduced at $6/month (¥39), 4.1B Credits https://mimo.mi.com/docs/en-US/tokenplan/Token%20Plan/subscription
2026-09-27 Token Plan Standard Baseline: introduced at $16/month (¥99), 11B Credits https://mimo.mi.com/docs/en-US/tokenplan/Token%20Plan/subscription
2026-09-27 Token Plan Pro Baseline: introduced at $50/month (¥329), 38B Credits https://mimo.mi.com/docs/en-US/tokenplan/Token%20Plan/subscription
2026-09-27 Token Plan Max Baseline: introduced at $100/month (¥659), 82B Credits https://mimo.mi.com/docs/en-US/tokenplan/Token%20Plan/subscription
2026-09-27 Pay-as-you-go (international) Baseline: mimo-v2.6-pro $0.435/$0.87 per M in/out, mimo-v2.6-flash $0.14/$0.28 https://mimo.mi.com/docs/en-US/price/pay-as-you-go

Compared to
#

  • OpenCode: the upstream project; choose OpenCode for the longer track record, the larger community, and the leanest measured token baseline, and MiMo Code only if the memory and long-horizon additions justify a fork you now depend on.
  • Kimi Code: the closest peer, another terminal harness tuned by a challenger lab for its own cheap models; pick on which vendor’s models and quota structure you prefer, not on benchmark claims.
  • Claude Code: the platform benchmark MiMo Code positions against; Claude Code buys polish and subscription simplicity, MiMo Code buys open source and low per-token cost.

Bottom line
#

Recommended for engineers who want an auditable terminal agent with explicit long-horizon state management and some of the cheapest credible token prices, and who are willing to run a 0.1.x tool. Not for teams that need release stability, and not because a vendor benchmark says it beats Claude Code past 200 steps. My disagreeable claim: the co-evolution story is mostly packaging, since the mature harness arrived with the OpenCode fork and the observable Xiaomi contribution is a checkpoint-and-memory layer plus a model bundle, which is genuinely useful but narrower than the tagline suggests.

Changes
#

  • 2026-09-27 - Created.
  • 2026-09-27 - Corrected the latest-release fact: the newest release by publication time is v0.1.14 (published 2026-09-23), a day after v0.1.15 (2026-09-22); the 0.1.x status is unchanged.

See also
#

  • OpenCode - the upstream fork this is built on, and the token-efficiency baseline it inherits
  • Kimi Code - the other challenger-vendor terminal harness tuned for its maker’s own cheap models
  • Claude Code - the harness the vendor benchmarks compare against, and the endurance-gap framing
  • Model selection for coding tasks - where MiMo’s price and long-horizon claims fit the model side of the decision
  • Agentic Coding Tools Landscape - the map this note’s category belongs to

References
#