MiMo-V2.6 for Coding Agents: What Changed and How to Run It

Published September 25, 2026 · by the AQ team

MiMo-V2.6 is Xiaomi's new family of open-weights models, released September 22, 2026. The flagship, MiMo-V2.6-Pro, is a 1.02-trillion-parameter sparse mixture-of-experts model with 42B active parameters, a 1M-token context window, natively omnimodal input (text, image, video, and audio), and an MIT license, and it debuted as the top open-weights model on the Artificial Analysis Intelligence Index with a score of 46 (Xiaomi cites 46.32). It ships alongside MiMo-V2.6-Flash, a smaller 309B-parameter sibling with 15B active, and a faster-serving Pro-UltraSpeed variant. For teams running coding agents, the headline is the combination: frontier-adjacent agentic coding scores, weights you can download, and list prices well under a dollar per million tokens.

What Xiaomi shipped

Specs and list prices below are from Xiaomi's model pages and OpenRouter, as of September 2026. A mixture-of-experts (MoE) model activates only a fraction of its parameters per token, which is how a 1T-parameter model serves at these prices:

Both open-weights models take text, images, video, and audio as input and produce text, and both support tool calling, which is the capability that actually matters for agent harnesses. Pro-UltraSpeed is the same model served much faster (Xiaomi claims up to 20 times faster output at the same quality) at roughly ten times the price, aimed at latency-sensitive interactive work. Cached input on Xiaomi's API is nearly free: $0.0036 per million tokens for Pro, which matters for agent loops that resend a large, stable prompt prefix every turn.

What Xiaomi says about coding

Xiaomi's release notes lead with agentic coding. On its published numbers, MiMo-V2.6-Pro scores 78.6 on SWE-Bench Verified and 34.9 on Terminal-Bench 4.0, and on DeepSWE v1.1 it reports 71.9, just behind the Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) scores it published alongside. The release notes also claim tool-calling accuracy in thinking mode rose from 64% to 97.0% versus the prior generation, which, if it holds up, addresses the most common failure mode of open models inside agent harnesses: not weak reasoning, but malformed or misdirected tool calls that stall a session.

The training story is as notable as the scores. Per launch coverage of Xiaomi's technical report, the final reinforcement-learning run took under six days and cost about $2.62 million for Pro (roughly $850,000 for Flash), covering about 30 RL steps over roughly 750,000 trajectories. Whatever else is true, the marginal cost of turning a strong base model into a strong agentic model has fallen far enough that a phone maker did it in a week, and shipped the result under MIT.

The independent read, launch week

Artificial Analysis measured MiMo-V2.6-Pro at launch and scored it 46 on its Intelligence Index, the highest score an open-weights model has posted, edging past the DeepSeek and GLM releases that traded that crown through 2026 (our DeepSeek V4 Pro guide covers the previous holder). That ties xAI's Grok 4.7, released the day before, while the closed frontier leaders sit at 53 on the same index as of late September 2026. Artificial Analysis also measured cost to run its full index at $0.13 per task, which puts the model on its intelligence-versus-cost Pareto frontier: nothing tracked is both smarter and cheaper per completed task.

The usual caveats apply, and one unusual one. Usual: vendor coding numbers and independent ones routinely diverge, because harness, tool configuration, and token budgets move agentic scores a lot; our guide to benchmarking the model or the harness covers why a single Terminal-Bench or SWE-Bench figure is a range, not a fact. Unusual: on September 10, 2026, Anthropic published a threat report accusing several Chinese labs, Xiaomi among them, of training on Claude outputs at industrial scale, citing more than 400,000 Claude requests across more than 1,500 accounts over 20 days in spring 2026, routed through proxy services. Xiaomi had not publicly responded as of September 25, 2026. The accusation does not change what the released weights measurably do, but teams with policies about training-data provenance should know the dispute exists before standardizing on the model.

How to run it in the agent CLIs

Xiaomi's API lives at https://api.xiaomimimo.com/v1 and speaks both the OpenAI and Anthropic wire protocols, with lowercase model IDs: mimo-v2.6-pro, mimo-v2.6-flash, and mimo-v2.6-pro-ultraspeed. The models are also on OpenRouter at the same list prices. That means any harness with a custom-provider or base-URL setting can drive it. Four that document the path, verified on each vendor's docs as of September 2026:

OpenCode. Custom providers are first-class: add a provider block to opencode.json pointing the OpenAI-compatible SDK at the endpoint, then list the models you want in the picker.

{
  "provider": {
    "xiaomi": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Xiaomi MiMo",
      "options": { "baseURL": "https://api.xiaomimimo.com/v1" },
      "models": { "mimo-v2.6-pro": { "name": "MiMo-V2.6-Pro" } }
    }
  }
}

Codex CLI. Codex reads custom providers from config.toml: define the provider with a base URL and an environment variable for the key, then select it. Keep the base URL without a trailing slash, and note that a project-level .codex/config.toml only loads once the project is trusted.

model = "mimo-v2.6-pro"
model_provider = "xiaomi"

[model_providers.xiaomi]
name = "Xiaomi MiMo"
base_url = "https://api.xiaomimimo.com/v1"
env_key = "MIMO_API_KEY"
wire_api = "chat"

Kimi Code CLI. Providers are configured in its config file under a [providers.name] table with type, base_url, and api_key; base URLs are set at the provider level, not per model, and OPENAI_API_BASE overrides the base URL for OpenAI-compatible providers.

Cursor. In Settings, then Models, add your key under the OpenAI API key section and enable the override for the OpenAI base URL. The override is global (it applies to every custom model you add, so you cannot mix two custom endpoints), the endpoint must speak the OpenAI chat-completions protocol, and Cursor may reject localhost or private-network URLs.

Self-hosting is the other route: the Pro and Flash checkpoints deploy on vLLM or SGLang, and the MIT license permits commercial use without a separate agreement. Our self-hosted agents guide covers what that actually takes at 1T-parameter scale.

When to pick it, when not to

Pick MiMo-V2.6 when price per completed task dominates: at $0.435 input and $0.87 output for Pro, an overnight batch of agent runs costs a fraction of frontier list prices, and Flash at $0.14 and $0.28 makes always-on background agents nearly free to keep busy. Pick it when you need open weights for audit, jurisdiction, or self-hosting reasons and want the strongest option in that class this month. And pick Flash specifically for high-volume, lower-stakes work: summarizing sessions, drafting tests, first-pass triage.

Look elsewhere when you need the strongest available model regardless of cost (the closed frontier still clears it by about seven points on the independent index, and by more on the hardest terminal-driving work), when your organization's provenance policies make the unresolved distillation dispute disqualifying, or when your workload is latency-sensitive and you are not willing to pay the UltraSpeed premium. And as with every launch week: the number that decides a default is not a leaderboard row, it is the same backlog task run on your own repository against your incumbent, judged by your own tests.

Where AQ fits

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. That makes a release like MiMo-V2.6 a same-day experiment instead of a migration: open two workspaces on the same repository (each is an isolated git worktree on its own branch), point OpenCode at Xiaomi's endpoint in one and run your incumbent in the other, give both the identical task, and let the whole team watch both live sessions from the browser. Sessions run in persistent tmux on your team's VM, so a long evaluation run survives closed laptops and resumes from any device, and everyone signs into the CLIs with their own accounts and keys, which AQ never marks up. When the comparison is done, our guide to reviewing an AI coding session covers what to look at beyond the final diff: how each model explored, what it verified, and where it guessed.