Overview
Xiaomi MiMo Open Platform is the official large model API platform launched by Xiaomi on 2026-03-18, uniformly hosting the full MiMo-V2.5 series (Pro / Flash / Omni / TTS / ASR). On April 28, Xiaomi announced that its flagship large model MiMo-V2.5-Pro and the full-modal MiMo-V2.5 series are fully open-sourced (MIT license), making it the first mobile phone manufacturer to top the global open-source large model leaderboard. On May 27, the API permanently reduced prices by 99% (cache hit input), simultaneously launching the Orbit incentive plan with 100 trillion tokens.
The platform has since moved to a new generation. On 2026-09-22, Xiaomi released and open-sourced the MiMo-V2.6 series in Pro and Flash variants — two native omni-modal models framed as a step toward recursive self-improvement. Pro is a sparse MoE at 1.02T total and 42B active parameters, Flash at 309B total and 15B active, both with a 1M-token context and text, image, video and audio input under MIT-licensed weights. Pro scored 46 on the Artificial Analysis Intelligence Index, the highest open-weight result, and measured 1,628 on the Code Arena WebDev arena. Training ran large-scale asynchronous GRPO in under six days at a reported 2.62 million and 850,000 dollars, with over 7,000 RL task environments, the training framework and Distill-Qwen-9B released alongside. API pricing carries over from the V2.5 series, and the MiMo Desktop client and membership plan launched at the same time. On 2026-09-24, Xiaomi detailed HySparse2, the core architecture for the next-generation MiMo-V3, cutting prefill compute by 5.02x and shrinking the KV cache by 4.5x at a 1-million-token length; in the same period MiMo lead Luo Fuli publicly streamed the MiMo-V2.6 reinforcement-learning post-training run over six days, at a total RL post-training cost above 3.47 million USD across roughly 750,000 trajectories.
Earlier milestones that shaped its presence: On 2026-06-11, Luo Fuli's team open-sourced the terminal programming Agent「MiMo Code」 (developed by 5 people in 14 days, based on OpenCode, topping GitHub Trending on the first day of open-source, with Stars quickly exceeding 9000+), positioning MiMo-V2.5-Pro as the 'domestic open-source version of Claude Code'; On 2026-06-09, in collaboration with TileRT, MiMo-V2.5-Pro-UltraSpeed was launched, the world's first model to achieve a decoding speed of 1000+ tokens/s at the trillion-parameter scale. Important reminder: The old versions MiMo-V2-Pro/Omni were taken offline on 2026-06-30, with automatic routing migration to MiMo-V2.5-Pro/Omni (using V2.5 pricing), and MiMo-V2-TTS is simultaneously routed to V2.5-TTS. Developers need to pay attention to interface migration.
Key Features
- Full MIT License Open Source: 8 models including Pro/Omni/TTS/ASR have full weights open-sourced on Hugging Face, supporting commercial use and secondary training; the MiMo-V2.6 Pro and Flash variants are likewise released under MIT-licensed weights
- Token Efficiency +42%: Compared to similar models, the number of tokens required for the same output is nearly halved, directly cutting Agent task costs in half
- First to Adapt 7 Domestic Chips: Covers almost all domestic inference chip platforms, friendly for domestic substitution
- Pure Browser Architecture: MiMo-V2.5 series natively designed for browser use, no App binding required
- 100 Trillion Token Incentive Plan: Orbit plan provides free computing power to developers to support innovation
- MiMo Code Terminal Programming Agent: Open-sourced on 2026-06-11, built by 5 people in 14 days, based on OpenCode, with MiMo-V2.5-Pro outperforming Claude Code+Sonnet 4.6 on three benchmarks
- UltraSpeed Extreme Inference: Released on 2026-06-09 in collaboration with TileRT, achieving a decoding speed of 1000+ tokens/s at the trillion-parameter scale, the world's first model at this scale and speed
- MiMo-V2.6 Pro / Flash (New): Released and open-sourced on 2026-09-22 as two native omni-modal models: Pro at 1.02T total and 42B active parameters, Flash at 309B total and 15B active, both with 1M-token context and text, image, video and audio input under MIT-licensed weights; Pro scored 46 on the Artificial Analysis Intelligence Index, the highest open-weight result, and measured 1,628 on the Code Arena WebDev arena
- HySparse2 Architecture for MiMo-V3 (New): Detailed on 2026-09-24 as the core architecture for the next-generation MiMo-V3, cutting prefill compute by 5.02x and shrinking the KV cache by 4.5x at a 1-million-token length
- MiMo Desktop Client and Membership (New): Launched on 2026-09-22 alongside the MiMo-V2.6 series
Use Cases
- Agent application development requiring long context and high cost-effectiveness
- Domestic large model substitution (finance, government, education, etc.)
- Pairing with third-party Agent frameworks like Claude Code/OpenClaw for programming workflows
- Multimodal applications (unified processing of images, video, audio, text)
Pros
- MIT license is the most permissive open-source license on the market, with zero barriers to commercial use
- After a permanent 99% price reduction, it is one of the cheapest flagship models globally
- Most complete adaptation for domestic chips, hassle-free local deployment
- Lei Jun's commitment of 60 billion AI investment over three years ensures long-term sustainability
- The MiMo-V2.6 generation extends the open-weight line with 1M-token context and native omni-modal input, and Pro's 46 on the Artificial Analysis Intelligence Index is the highest open-weight result
- HySparse2 lays out the architecture path for the next-generation MiMo-V3, cutting prefill compute by 5.02x and shrinking the KV cache by 4.5x at a 1-million-token length
Pricing
Pro ≤256K: Input $1.00/M, Output $3.00/M; Pro ≤1M: Input $2.00/M, Output $6.00/M. Omni ≤256K: Input $0.40/M, Output $2.00/M. TTS is free for a limited time. New users receive ¥10 trial credit upon registration, and participating in the Orbit plan can earn millions of token quotas. Note: Old versions MiMo-V2-Pro/Omni were taken offline on 2026-06-30, uniformly routed under V2.5 series pricing. The MiMo-V2.6 series carries API pricing over from the V2.5 series, and the MiMo Desktop client and membership plan launched alongside it on 2026-09-22.
Summary
Xiaomi MiMo is the most noteworthy new force in the 2026 domestic open-source large model camp—'topping open-source + MIT license + permanent price reduction + domestic chip adaptation' all done together, with a very clear positioning. In June it gained attention twice with MiMo Code (terminal programming Agent) and UltraSpeed (1000+ tokens/s extreme inference), and in September it advanced to the MiMo-V2.6 generation — Pro and Flash native omni-modal models with 1M-token context, MIT-licensed weights, and Pro's 46 on the Artificial Analysis Intelligence Index as the highest open-weight result — while detailing HySparse2 as the core architecture for the next-generation MiMo-V3. Its differentiation label of 'cost-effectiveness + extreme engineering' now sits alongside a clear roadmap toward recursive self-improvement. Suitable for budget-sensitive Agent developers, enterprises needing domestic solutions, and teams wanting a programming flagship but put off by Claude Opus's price.
Version History
- HySparse2 unveiled as the core architecture for MiMo-V3 (2026-09-24): Xiaomi detailed HySparse2, the core architecture for the next-generation MiMo-V3, cutting prefill compute by 5.02x and shrinking the KV cache by 4.5x at a 1-million-token length. In the same period MiMo lead Luo Fuli publicly streamed the MiMo-V2.6 reinforcement-learning post-training run over six days, at a total RL post-training cost above 3.47 million USD across roughly 750,000 trajectories.
- Xiaomi releases and open-sources MiMo-V2.6 in Pro and Flash variants (2026-09-22): Xiaomi released and open-sourced the Xiaomi MiMo-V2.6 series, comprising two native omni-modal models, Pro and Flash, which the company frames as a step toward recursive self-improvement. Pro is a sparse MoE at 1.02T total and 42B active parameters, Flash at 309B total and 15B active, both with a 1M-token context and text, image, video and audio input under MIT-licensed weights. Pro scored 46 on the Artificial Analysis Intelligence Index, the highest open-weight result, and measured 1,628 on the Code Arena WebDev arena. Training ran large-scale asynchronous GRPO in under six days at a reported 2.62 million and 850,000 dollars, with over 7,000 RL task environments, the training framework and Distill-Qwen-9B released alongside. API pricing carries over from the V2.5 series, and the MiMo Desktop client and membership plan launched at the same time.
- MiMo-V2-Pro/Omni taken offline, full migration to V2.5 (2026-06-30): Old model services stopped, automatically routed to MiMo-V2.5-Pro/Omni (using V2.5 pricing), MiMo-V2-TTS simultaneously routed to V2.5-TTS
- MiMo Code open-sourced (2026-06-11): Luo Fuli's team of 5 built the terminal programming Agent in 14 days, based on OpenCode, topping GitHub Trending on the first day of open-source, with Stars quickly exceeding 9000+
- MiMo-V2.5-Pro-UltraSpeed (2026-06-09): Released in collaboration with TileRT, the world's first model to achieve a decoding speed of 1000+ tokens/s at the trillion-parameter scale