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

Updated November 4, 2026 · first published September 25, 2026 · by the AQ team

To deploy MiMo for coding agents, connect Xiaomi's API to a CLI with a custom provider, install Xiaomi's own MiMo Code CLI, or serve the MIT-licensed weights on vLLM or SGLang. The API starts at $0.14 input and $0.28 output per million tokens for Flash; MiMo Code's source is free, with model usage billed through the account or provider you connect; self-hosting means paying for your own compute and operations. Use the API to try MiMo in your existing workflow, MiMo Code for Xiaomi's terminal agent, and self-hosting when you need control over inference infrastructure.

Deploy MiMo for coding agents: three routes

The execution machine and the inference endpoint are separate choices. A CLI can edit a local repository while sending prompts and tool results to a hosted model. The session visibility column below describes that data flow, not a promise about retention.

Xiaomi describes its service as an OpenAI and Anthropic compatible gateway. Its API guide gives https://api.xiaomimimo.com/v1 for OpenAI-compatible pay-as-you-go requests and https://api.xiaomimimo.com/anthropic for Anthropic Messages. Token Plan subscriptions use their own endpoint and credentials, so use the pair shown in your console.

Current models and API prices

These are Xiaomi's overseas real-time API rates, checked November 4, 2026, per million tokens in US dollars. Xiaomi's home page lists the prices; its pricing documentation distinguishes real-time and batch billing.

The model documentation lists a 1M-token context window, 128K maximum output, multimodal understanding, text generation and function calling. The Pro model card specifies 1.02T total parameters with 42B activated per token; the Flash card specifies 309B total with 15B activated. Both accept text, image, video and audio input. Xiaomi advertises UltraSpeed as Pro quality with up to 20 times faster output; that is a vendor claim, and its token rates are ten times Pro's.

Batch API: Xiaomi now documents overseas batch rates for Pro of $0.2175 input, $0.0018 cache hit and $0.435 output per million tokens. Flash batch rates are $0.07, $0.0014 and $0.14 respectively. UltraSpeed does not support batch. This is asynchronous inference for evaluations and other work that can wait; an interactive coding loop still needs tool results between turns. Obtain the dedicated batch base URL from the console.

MiMo Code: Xiaomi's terminal coding agent

According to the MiMo Code README, MiMo Code is a fork of OpenCode that reads and writes code, runs commands and manages Git. It supports MiMo-V2.6-Pro and Flash, plus custom OpenAI-compatible providers. Install it on macOS or Linux, then run it in your project:

curl -fsSL https://mimo.xiaomi.com/install | bash
cd your-project
mimo

On Windows, use the README's PowerShell installer:

powershell -ep Bypass -c "irm https://mimo.xiaomi.com/install.ps1 | iex"
mimo

First launch guides you through authentication and provider selection. The README documents MCP connections, parallel subagents, persistent project memory, mimo run for headless execution and mimo serve for a remote server that a local terminal can attach to. Custom Provider in the terminal UI connects another OpenAI-compatible API.

The CLI source has an MIT license, and the README separately says use is subject to USE_RESTRICTIONS.md; Xiaomi-hosted services have their own terms. Review those alongside the source license. We installed the CLI without signing in on November 4, 2026: mimo --version returned 0.1.16.

Connect Xiaomi's API to an existing agent

The examples below use pay-as-you-go credentials and were checked against vendor documentation on November 4, 2026. Set MIMO_API_KEY in the environment of the CLI process. Keep real keys out of repository files.

OpenCode

OpenCode documents custom providers with @ai-sdk/openai-compatible for Chat Completions. Add this to opencode.json, then choose xiaomi/mimo-v2.6-pro from /models. Change the model ID to mimo-v2.6-flash to evaluate Flash.

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

Codex CLI

The current Codex configuration reference supports only wire_api = "responses". Xiaomi's Codex guide documents that protocol and supplies a model catalog file. Download its linked model-catalogs.json to ~/.codex/model-catalogs.json, then configure ~/.codex/config.toml:

model_catalog_json = "~/.codex/model-catalogs.json"
model = "mimo-v2.6-pro"
model_provider = "xiaomi"
model_reasoning_effort = "high"
model_supports_reasoning_summaries = true
model_reasoning_summary = "none"
model_context_window = 1048576
web_search = "disabled"

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

This uses Codex's documented environment-key field instead of storing a token in TOML. Xiaomi says an error about custom tools requiring MiMo freeform Responses lite mode calls for its model catalog configuration and a Codex restart.

Kimi Code CLI

The current Kimi Code configuration docs use ~/.kimi-code/config.toml and provider type openai for Chat Completions. They also document api_key_env as an explicit way to read a shell variable. Add a provider and model alias:

default_model = "mimo-pro"

[providers.xiaomi]
type = "openai"
base_url = "https://api.xiaomimimo.com/v1"
api_key_env = "MIMO_API_KEY"

[models.mimo-pro]
provider = "xiaomi"
model = "mimo-v2.6-pro"
max_context_size = 1048576
capabilities = ["tool_use"]

This example declares tool use for coding. See Kimi's provider reference for additional model capabilities and its credential rules before adapting an older config: ordinary shell exports do not automatically override provider credentials or base URLs.

Cursor

Cursor's current API-key documentation describes keys for listed providers, but does not establish a supported MiMo custom-endpoint setup. It also says requests pass through Cursor's servers for prompt building and that custom keys work only with chat models. We cannot verify the earlier base-URL recipe from those docs, so use one of the documented CLI routes above for a repeatable MiMo evaluation.

Self-host MiMo on vLLM or SGLang

The current Pro and Flash MOPD cards label the weights MIT and provide commands for both serving engines. Pro's vLLM example uses tensor parallel size 8; Flash's uses 4. Pro's SGLang example specifies two nodes with tensor parallel size 16, while Flash's specifies tensor parallel size 8. Both cards configure MiMo reasoning and tool-call parsers. Follow the linked recipes and check compatibility with your serving version before provisioning.

The current vLLM Pro recipe gives a 566 GB checkpoint and hardware prerequisites of 8 H200 GPUs, 4 MI355X GPUs, or equivalent aggregate VRAM of at least 680 GB. The Flash recipe gives a 173 GB checkpoint and prerequisites of 4 H200 GPUs, 4 MI355X/MI325X/MI300X GPUs, or at least 208 GB aggregate VRAM. These are the recipes' stated configurations, not a guarantee for every context length or concurrency level.

Those vLLM recipes require version 0.31.0 or later to load the stored weights and recommend vllm/vllm-openai:v0.31.0 over the older V2.5 image still linked from the cards. Check the recipe's hardware-specific instructions, including newer nightly requirements for Flash on MI300X and MI325X. An hourly hosting cost is not documented. Pro's 42B active parameters describe computation per token; the checkpoint still has 1.02T total parameters. Our self-hosted AI coding agents guide covers infrastructure ownership, network access and the operational work around a model server.

Use the Pro deployment instructions or Flash deployment instructions, then point the CLI's custom provider at your server. The cards also include a use notice that explicitly says it does not amend the model license or add license conditions.

Updated checkpoints and the V2.5 migration

Xiaomi's September 27 technical post explains a failure in which Pro and Flash repeated similar tool calls without progress. The MOPD cards confirm the updated checkpoints mitigate that behavior. Xiaomi says the updated models reached its API on September 25, 2026 at 06:00 UTC+8, retaining the IDs mimo-v2.6-pro and mimo-v2.6-flash. Self-hosters should distinguish the -RL and -MOPD checkpoint names when recording results.

V2.5 IDs no longer work. Xiaomi's retirement notice gives October 21, 2026 at 10:00 UTC+8 as the cutoff. Replace mimo-v2.5-pro with mimo-v2.6-pro and mimo-v2.5 with mimo-v2.6-flash in provider configs, scripts and saved job definitions.

The notice says thinking mode fixes temperature at 1.0 and top_p at 0.95. Flash's default max_completion_tokens is 131072, up from V2.5's 32768. Set an explicit output budget appropriate to your task and test the migrated configuration before resuming scheduled runs.

Coding benchmarks and cost per completed task

As checked November 4, 2026, Xiaomi's release page reports Pro at 71.9 on DeepSWE v1.1 and 34.9 on Terminal Bench 4.0, with Flash at 67.9 and 28.8. These are vendor-reported results. They suggest Pro is a candidate for difficult repository work and Flash for a lower-cost first pass; they do not establish which will finish your backlog tasks reliably.

Separately, Artificial Analysis's Pro page, read November 4, 2026, shows an Intelligence Index score of 30 and a cost of $0.26 per index task. That is a broad independent evaluation, not a coding-only score or a forecast of your agent bill. The current figures replace the September launch snapshot in this guide; do not compare index snapshots without checking the evaluation version.

Harness, tools, context handling and token budgets can change agent results. Our guide to benchmarking the model or the harness explains how to control those variables. Run the same repository task with Pro, Flash and your incumbent, then record accepted changes, test results, retries, elapsed time and total spend. Divide spend by tasks that meet your acceptance criteria: cheap tokens can still produce an expensive task if the agent needs repeated attempts.

There is also a dated provenance dispute. Anthropic's September 2026 threat report accuses Xiaomi of distilling Claude outputs, citing more than 400,000 requests across more than 1,500 accounts over 20 days in March and April 2026. This is Anthropic's allegation. Xiaomi had not publicly responded as of September 25, 2026; we found no later first-party response in this review. Teams with training-data provenance requirements should review the report before adopting the model.

Where AQ fits

AQ is the multiplayer coding harness where teams chat, build and run bots with every AI coding agent, like Claude Code and Codex: shared live terminals, review and merge, app previews and standing jobs, in your own cloud.

As described in AQ's product summary and What is AQ?, teams run agent CLIs, including OpenCode, in isolated git worktrees on a team's machine or in the customer's cloud. Point OpenCode at Xiaomi's API in one workspace and run your incumbent in another on the same repository. Each gets its own branch, and teammates can watch both live sessions and review the results together.

AQ never marks up usage on your own subscriptions or keys. Its built-in agent also runs on AQ models at the per-token price shown before you pick one; see AQ pricing and models and Auto. For the MiMo comparison above, the model connection belongs to OpenCode's custom provider. Use our session review guide to compare how each run explored the repository and verified its changes.