Published Updated
In one or two sentences, before anything else.
Xiaomi MiMo is Xiaomi's AI model family and platform: chat in MiMo AI Studio, an OpenAI- and Anthropic-compatible API, voice models, and coding and desktop agents. Its MiMo-V2.6-Pro, released with MIT-licensed weights on September 22, 2026, is the top open-weights model on Artificial Analysis's index as of October 17, 2026.
What it can do, one sentence each, and whether we actually tested it.
Two models that read text, images, video and audio in one context of up to 1M tokens and write up to 128K tokens.
Per Xiaomi's model list and model cards: Pro has 1.02T parameters with 42B active, Flash 309B with 15B active.
Pro and Flash weights, the technical report, about 7,000 RL task environments and the RL training code are public, with commercial use allowed.
eWEEK notes the extra 9B checkpoint is built on Alibaba's Qwen3.5-9B, so its license chain needs its own check.
MiMo-V2.6-Pro-UltraSpeed, which Xiaomi says writes up to 20 times faster than Pro at the same quality, for latency-sensitive work.
It costs ten times the Pro price per token.
OpenAI- and Anthropic-compatible endpoints with web search, structured output, function calling and a batch API, documented for Claude Code, OpenCode, Codex, OpenClaw and Cline.
Per Xiaomi's integration guides.
MiMo-V2.5-TTS with preset voices, voices designed from a text description and voices cloned from samples; MiMo-V2.5-ASR for Chinese, English and dialects such as Cantonese and Wu.
Text-to-speech is free for a limited time.
Browser chat with the V2.6 models, image and multimodal understanding, and multi-round web search added on October 15, 2026.
An MIT-licensed terminal coding agent built on OpenCode, with persistent memory, a plan-then-build Compose mode and voice input; it also accepts DeepSeek, Kimi and GLM keys.
Desktop reads files in many formats, drives the browser and, outside China, the computer; Claw is a cloud OpenClaw agent with Kingsoft Office documents.
Xiaomi lists Claw at 19.9 yuan a month and says overseas subscriptions are not available yet.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
Artificial Analysis's standard evaluation suite, sent to Xiaomi's own API endpoint.
What happened
It scored 46, the top open-weights result on the leaderboard, at $0.13 per index task against $5.41 for Qwen3.8 Max (0902). The catch is speed: 48 tokens per second, below the 68.2 median for comparable open models.
Case 2
Input
Users' own tasks in Arena's agent mode, 27,455 sessions with MiMo-V2.6-Pro.
What happened
Arena measured a net improvement of +3.95% (±1.09), the best of the open-weights models and ahead of DeepSeek-V4.1-Flash (Max) at +2.74%. Closed models such as Claude Opus 5.5 (Max), at +18.88%, remain far ahead.
Case 3
Input
Set ANTHROPIC_BASE_URL to https://api.xiaomimimo.com/anthropic or the Token Plan endpoint, put a MiMo key in ANTHROPIC_AUTH_TOKEN, set every model slot to mimo-v2.6-pro and add [1m] for the 1M context.
What happened
Xiaomi documents the same pattern for OpenCode, Codex, OpenClaw and Cline. This is a documented setup, not a tested result, and nobody outside Xiaomi has repeated its claim that MiMo Code beats Claude Code on SWE-Bench Pro with the same model.
Case 4
Input
Formalize the main theorem of Li and Yorke's paper Period Three Implies Chaos in Lean 4, using subagents under a strategy the researchers designed.
What happened
Xiaomi says the project grew past 6,000 lines and passed Lean's kernel check with no unfinished placeholders, without Lean-specific training. Researchers guided and revised the work, and no one outside Xiaomi has reproduced it.
What it does badly, with the evidence, and what we could not settle.
Weak
Two model names are about to stop working.
Evidence: Xiaomi retires mimo-v2.5-pro and mimo-v2.5 at 10:00 Beijing time on October 21, 2026, and since October 14 it routes them to V2.6, which fixes temperature at 1.0 in thinking mode and raises the default output length. Anyone pinned to V2.5 who cannot retest this week should not depend on MiMo right now.
Behind closed frontier models on long terminal and security work.
Evidence: Xiaomi's own table, reported by 36Kr on September 22, 2026, gives MiMo-V2.6-Pro 34.9 on Terminal Bench 4.0 against 59.6 for GPT-6 Astra, with lower results than leading closed models on several cybersecurity tests.
Slow output unless you pay for UltraSpeed.
Evidence: Artificial Analysis measured 48 tokens per second on Xiaomi's API against a 68.2 median for comparable open-weights models; UltraSpeed costs ten times the Pro price.
Hard to run yourself.
Evidence: The MiMo-V2.6-Pro-RL weights total about 566 GB on Hugging Face for a 1.02T-parameter model, served through SGLang or vLLM per the model card.
Uncertain
How good it is beyond the benchmarks Xiaomi chose.
Why unsure: Most published scores are Xiaomi's. The independent ones are Artificial Analysis and Arena, and the Arena text score rests on only 4,075 votes; eWEEK notes Epoch AI considers SWE-bench Verified and DeepSWE flawed.
What happens to your data.
Why unsure: The platform documentation we read has no statement on whether API or AI Studio inputs are used for training, and we did not find a privacy policy there.
The exact switch-over time on October 14.
Why unsure: Xiaomi's deprecation page gives 10:00 (UTC+8) in its table and 18:00 (UTC+8) in the note under it.
The release day of MiMo-V2.6.
Why unsure: Xiaomi's release log and announcement say September 22, 2026, while Artificial Analysis lists September 21 and the Hugging Face repository was created on September 21 UTC.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Maker | Xiaomi, through its MiMo team | Source: mimo.xiaomi.comchecked |
| Current models | MiMo-V2.6-Pro, MiMo-V2.6-Flash and MiMo-V2.6-Pro-UltraSpeed (September 22, 2026); MiMo-V2.5-TTS series; MiMo-V2.5-ASR | Source: mimo.mi.comchecked |
| Context and inputs | 1M-token context, up to 128K output tokens; text, image, video and audio input | Source: mimo.mi.comchecked |
| Size and license | Pro: 1.02T parameters, 42B active; Flash: 309B, 15B active; both MIT | Source: huggingface.cochecked |
| API price, outside China | Pro $0.435 in / $0.87 out; Flash $0.14 / $0.28; UltraSpeed $4.35 / $8.70 per million tokens | Cache hits from $0.0028; cache writes free for a limited time. Source: mimo.mi.comchecked |
| API price, China | Pro ¥3 in / ¥6 out; Flash ¥1 / ¥2; UltraSpeed ¥30 / ¥60 per million tokens | Source: mimo.mi.comchecked |
| Voice pricing | TTS models free for a limited time; ASR $0.074 an hour (¥0.5 in China) | Source: mimo.mi.comchecked |
| Token Plan | Lite $6 (¥39), Standard $16 (¥99), Pro $50 (¥329), Max $100 (¥659) a month | Source: mimo.mi.comchecked |
| Retiring models | mimo-v2.5-pro and mimo-v2.5 stop at 10:00 on October 21, 2026 (UTC+8); routed to V2.6 since October 14 | Source: mimo.mi.comchecked |
| Independent ranking | Artificial Analysis Intelligence Index 46, the highest of 67 open-weights models listed | Source: artificialanalysis.aichecked |
Answer first, then why, an example, the limits and what to do next.
Yes, Xiaomi MiMo's current V2.6-Pro and V2.6-Flash models have open weights under the MIT license. Xiaomi put them on Hugging Face on September 22, 2026 with the technical report and RL code; the V2.5 series is MIT too, while MiMo-V2-Pro and V2-Omni were API-only.
Limits
Open weights do not make the 1.02T-parameter Pro model easy to run: its files total about 566 GB.
The MiMo API charges $0.435 per million input tokens and $0.87 per million output tokens for MiMo-V2.6-Pro. MiMo-V2.6-Flash costs $0.14 and $0.28, UltraSpeed $4.35 and $8.70; in China the same models cost ¥3 and ¥6, ¥1 and ¥2, and ¥30 and ¥60.
Limits
Free cache writes and free text-to-speech are limited-time offers, and web search is billed per call.
Yes, mimo-v2.5-pro and mimo-v2.5 stop working at 10:00 Beijing time on October 21, 2026. Since October 14 Xiaomi has routed requests for them to mimo-v2.6-pro and mimo-v2.6-flash; after October 21 the old names return an error.
Next step
Change the model name and retest, because thinking-mode temperature and the default output length differ in V2.6.
MiMo-V2.6-Pro is the strongest open-weights model on Artificial Analysis's index, but it does not lead every board. It scores 46 there against 45 for Alibaba's closed Qwen3.8 Max (0902) and 39 for DeepSeek V4.1 Flash; on Arena's text board Qwen3.8-Max scores 1482 to MiMo's 1479, and no public benchmark we found ranks it against ByteDance's current Doubao models.
Why
Artificial Analysis also puts MiMo-V2.6-Pro's cost per index task at $0.13, against $5.41 for Qwen3.8 Max (0902) and $0.27 for DeepSeek V4.1 Flash.
Xiaomi offers no official MiMo phone app that we could find. Its platform FAQ says it has not launched a standalone MiMo app and warns against downloads from unofficial channels; the official ways in are the browser-based MiMo AI Studio, MiMo Desktop and the MiMo Code terminal tool.
Yes, MiMo works in Claude Code through Xiaomi's Anthropic-compatible endpoint. Xiaomi's guide sets ANTHROPIC_BASE_URL to https://api.xiaomimimo.com/anthropic for pay-as-you-go or to the Token Plan endpoint, with a MiMo key and mimo-v2.6-pro as the model.
This page is built from Xiaomi's MiMo site, its platform documentation and price pages, the Hugging Face model cards, reporting by 36Kr and eWEEK, and three public rankings: Artificial Analysis, Arena's text leaderboard and Arena's Agent Arena. Scores that only Xiaomi has published are labelled as Xiaomi's. We checked everything on October 17, 2026.
MiMo is Xiaomi's own family of AI models and the services built on them. Users chat with the models in MiMo AI Studio in a browser; developers call them through the MiMo API Open Platform, which accepts OpenAI- and Anthropic-format requests. Around the models Xiaomi ships MiMo Code, an MIT-licensed terminal coding agent built on OpenCode (Xiaomi), MiMo Desktop, an agent app that left early access on September 22, 2026, and MiMo Claw, a cloud agent based on OpenClaw (Xiaomi).
| Model | Released (per Xiaomi) | Status on October 17, 2026 |
|---|---|---|
| MiMo-V2-Flash | December 16, 2025 | Open weights; retired from the API June 30, 2026 |
| MiMo-V2-Pro, V2-Omni, V2-TTS | March 18, 2026 | Released on the API; retired June 30, 2026 |
| MiMo-V2.5, V2.5-Pro, V2.5-TTS series | April 23, 2026 | V2.5 and V2.5-Pro weights under MIT; both retire October 21, 2026 |
| MiMo-V2.5-ASR | June 2, 2026 on the API | Open-source speech recognition |
| MiMo-V2.6-Pro, V2.6-Flash, V2.6-Pro-UltraSpeed | September 22, 2026 | Current; Pro and Flash weights under MIT |
Sources: Xiaomi's release log, deprecation page and V2.5 open-source notice.
Artificial Analysis gives MiMo-V2.6-Pro 46 on its Intelligence Index v4.3.2, the highest of the 67 open-weights models on its leaderboard; the top model overall, Claude Opus 5.5, scores 58. It measured 48 tokens per second on Xiaomi's API, below the 68.2 median of comparable open-weights models (Artificial Analysis).
On Arena's text leaderboard, updated October 15, 2026, MiMo-V2.6-Pro scores 1479 from 4,075 votes, level with GLM-5.3 (Max); the leader, Gemini 4 Argon (High), has 1524 (Arena). Arena's Agent Arena, updated October 16, gives it a net improvement of +3.95% (±1.09) over 27,455 real agent sessions, the best of the open-weights models listed. Its predecessor MiMo-V2.5-Pro sits at -8.01% (Arena).
Xiaomi's own tables, reported by 36Kr, show the weak spots: 34.9 on Terminal Bench 4.0 against 59.6 for GPT-6 Astra, and lower results than the leading closed models on several cybersecurity tests. eWEEK adds that Epoch AI treats SWE-bench Verified and DeepSWE, two benchmarks Xiaomi relies on, as flawed.
| Model | Input per million tokens | Output per million tokens |
|---|---|---|
| MiMo-V2.6-Flash | $0.14 (¥1 in China) | $0.28 (¥2) |
| MiMo-V2.6-Pro | $0.435 (¥3) | $0.87 (¥6) |
| MiMo-V2.6-Pro-UltraSpeed | $4.35 (¥30) | $8.70 (¥60) |
Cached input starts at $0.0028, the batch API halves the Pro and Flash prices, text-to-speech is free for a limited time and speech recognition costs $0.074 an hour (pricing). Token Plan subscriptions for coding tools run from $6 (¥39) to $100 (¥659) a month (Token Plan).
| Model | Weights | AA Intelligence Index | AA cost per index task | Arena text score |
|---|---|---|---|---|
| MiMo-V2.6-Pro | MIT | 46 | $0.13 | 1479 |
| Qwen3.8 Max (0902) | Closed | 45 | $5.41 | 1482 |
| DeepSeek V4.1 Flash (max) | MIT | 39 | $0.27 | 1475 |
| DeepSeek V4 Pro 0813 | MIT | 36 | $0.67 | 1465 |
| ByteDance Dola Seed 2.0 Pro | Closed | Not listed | Not listed | 1456 |
Sources: Artificial Analysis and Arena. MiMo-V2.6-Pro leads the open models on Artificial Analysis and is the cheapest per task there, while Alibaba's closed Qwen3.8 Max edges it on Arena's text board. DeepSeek's own price page lists V4.1 Flash at $0.30 in and $1.20 out per million at peak hours, half that off-peak (DeepSeek). We found no public benchmark that ranks MiMo against ByteDance's current Doubao models directly: Artificial Analysis lists only Doubao Seed Code.
Sources
Changelog
Duck.ai is DuckDuckGo's private AI chat, launched in June 2024 as DuckDuckGo AI Chat. It lets you use models from OpenAI, Anthropic, Mistral and others without an account, and DuckDuckGo strips your IP address before each request. Six models are free; stronger ones come with DuckDuckGo's 9.99 or 19.99 USD monthly plans.
Gemini Notebook is Google's research notebook that answers from the sources you add and turns them into audio, video, reports and quizzes. Google renamed it from NotebookLM on July 16, 2026; it is the same standalone product, existing notebooks carried over, and notebooklm.google.com now redirects to notebook.google.com.
Grok is the chatbot from SpaceXSI, the Musk company formerly called xAI and then SpaceXAI, sold since October 13, 2026 in the X Subscription. Since January 2026 it has faced country blocks, investigations and lawsuits over sexualized deepfakes, and Canada's Privacy Commissioner found in June 2026 that X Corp. and xAI broke federal privacy law.
Character.AI is a chatbot app where people talk to characters other users create, made by Character Technologies. The company filed for Chapter 11 bankruptcy on October 12, 2026, and Disney bid 20 million USD for a license to some assets; the app runs during the sale and will be wound down if no bid keeps it going.
VideoEnhancer is a web tool that upscales short MP4 clips to 1080p, 2K or 4K with the FlashVSR model and also compresses, face-swaps and generates video. After a 5-second free trial, 1080p costs 3 credits a second, $0.03 to $0.09 depending on how credits are bought. VideoEnhancer is our own product: it is made by the team behind GenGrades.
LayerGrab splits a flat image into up to 16 named, transparent layers over a filled-in background, on the web, in Figma or in Chrome, or pulls out only what you name. Each layer costs $0.10 on its $12 plan, $0.08 on Premium. No independent review exists yet. LayerGrab is our own product: it is made by the team behind GenGrades.