MiMo-V2.6
Xiaomi's omnimodal agent models: a 1T Pro, a 310B Flash and a 9B research checkpoint, 1M context, MIT-tagged.
In short. MiMo-V2.6 is Xiaomi's family of AI models that read text, images, video and audio, and are tuned for coding and agent work. The big Pro and Flash versions need a server, but Xiaomi also released a 9-billion-parameter version that a well-equipped laptop can run. Hugging Face lists all of them as MIT-licensed.
Summary
Xiaomi announced and open-sourced MiMo-V2.6 on 21 September 2026. The series has two natively omnimodal mixture-of-experts models. MiMo-V2.6-Pro has about 1.02 trillion parameters with 42 billion active per token; MiMo-V2.6-Flash has about 310 billion with 15 billion active. Both accept text, images, video and audio, write text, and support a 1-million-token context.
The checkpoints are compact for their size because the expert weights are stored in MXFP4 (4-bit) while the rest is FP8 and BF16. Pro takes about 573 GB and Flash about 178 GB. The launch checkpoints are named MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL. On 27 September Xiaomi added MOPD upgrades of both, which fuse several domain-specialized teachers and, according to the cards, mitigate a failure where the model keeps repeating the same tool call without making progress.
Xiaomi says the series was trained with one large mixed reinforcement-learning run across coding, general agents, visual tasks and cybersecurity. It reports 71.9 for Pro and 67.9 for Flash on DeepSWE v1.1; these figures are self-reported. It also released MiMo-V2.6-Distill-Qwen-9B, a supervised fine-tune of Qwen3.5-9B that is meant as a starting point for agent research, plus a set of RL training environments and a training framework.
Hugging Face tags every repository as MIT, and Xiaomi's announcement calls the release fully open source. However, none of the repositories contains a LICENSE file as of 8 November 2026, so if you build a product on it, ask Xiaomi to confirm the terms in writing.
Good for
- Coding and tool-using agents that also need to understand screenshots, video or audio
- Long repositories and multi-session agent runs, thanks to the 1M-token window
- Flash for teams with a single large-memory node; the 9B checkpoint for laptops and agent research
- Studying reinforcement-learning recipes, since Xiaomi also released environments and a training framework
Watch out for
- The repositories are tagged MIT but contain no LICENSE file; get written confirmation before shipping a commercial product
- Pro needs roughly 573 GB of memory even at its shipped 4-bit expert precision, and Flash about 178 GB
- The 9B model is a fine-tune of Qwen3.5-9B, which is Apache 2.0
- Benchmark scores are Xiaomi's own and have not been independently reproduced
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the V2.6 Distill Qwen 9B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| V2.6 Distill Qwen 9B BF16 | 9.4B | 22 GB | 11 GB | 6.49 GB | Laptop or phone |
| V2.6 Flash MXFP4 | 310B 15B active | 713 GB | 357 GB | 196 GB | Big-memory desktop |
| V2.6 Pro MXFP4 | 1.02T 42B active | 2.35 TB | 1.17 TB | 630 GB | GPU server |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- Best open-source models for coding and AI agents
- Best open-source multimodal models in 2026 (images, video and audio)
- Open models that listen: nine take audio, but only three answer in speech
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is MiMo-V2.6?
- MiMo-V2.6 is Xiaomi's family of AI models that read text, images, video and audio, and are tuned for coding and agent work. The big Pro and Flash versions need a server, but Xiaomi also released a 9-billion-parameter version that a well-equipped laptop can run. Hugging Face lists all of them as MIT-licensed.
- Can I use MiMo-V2.6 commercially?
- Yes. MiMo-V2.6 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is MiMo-V2.6 open source?
- MiMo-V2.6 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run MiMo-V2.6 on my own computer?
- Yes. The smallest variant, V2.6 Distill Qwen 9B, needs about 6.49 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was MiMo-V2.6 released?
- MiMo-V2.6 was released by Xiaomi on September 21, 2026.
- How long a context window does MiMo-V2.6 have?
- MiMo-V2.6 supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Xiaomi: Introducing the MiMo-V2.6 series mimo.xiaomi.com
- MiMo-V2.6-Pro-RL model card (Hugging Face) huggingface.co
- MiMo-V2.6 collection (Hugging Face) huggingface.co
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.8-Flash-Next Alibaba Qwen | 180B 6B active | 256K | Qwen Community 1.0 (Conditional) | 124 GB Big-memory desktop | |
| MiniMax M3.1 Flash MiniMax | 428B 23B active | 1M | MiniMax Community (Conditional) | 277 GB Big-memory desktop | |
| GLM-5.3-Flash Z.ai (Zhipu AI) | 320B 18B active | 1M | MIT (Permissive) | 221 GB Big-memory desktop | |
| Xiaomi-OCR-0 Xiaomi | 873M | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone |