What is MiMo-V2.6-Pro?
MiMo-V2.6-Pro is Xiaomi's flagship language model, released with open weights under the MIT licence on 21 September 2026 alongside a smaller MiMo-V2.6-Flash. It is a sparse mixture-of-experts model with 1.02 trillion parameters, 42 billion active per token, and a 1-million-token context. It reads text, images, video and audio and writes text, using its own 681-million-parameter vision encoder and a separate audio encoder.
The Flash sibling has 309 billion parameters, 15 billion active, with the same context and inputs. Both are on Hugging Face and ModelScope, in Xiaomi's MiMo Studio, MiMo Code and MiMo Desktop apps, on the Xiaomi MiMo API platform and on OpenRouter. On 27 September Xiaomi added MOPD checkpoints of both, retrained to cut a habit of repeating the same tool call in agent loops.
On Xiaomi's API, Pro costs $0.435 per million input tokens and $0.87 per million output tokens, with cached input at $0.004; Flash costs $0.14 and $0.28. An UltraSpeed version of Pro runs about 20 times faster at ten times the price.
How much does MiMo-V2.6-Pro cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | MIT licence |
| MiMo-V2.6-Pro | $0.435 in / $0.87 out per 1M tokens | Cached input $0.004 |
| MiMo-V2.6-Flash | $0.14 in / $0.28 out per 1M tokens | |
| Pro UltraSpeed | $4.35 in / $8.70 out per 1M tokens | About 20 times faster output |
Open weights under MIT. Xiaomi API: Pro $0.435 input / $0.87 output per 1M tokens, cached input $0.004; Flash $0.14 / $0.28; Pro UltraSpeed $4.35 / $8.70; Batch API half price.
Is MiMo-V2.6-Pro worth it?
MiMo-V2.6-Pro is the strongest open-weight model available. Artificial Analysis scores it 46 on its Intelligence Index, first of 121 open models of its size, ahead of GLM-5.3 at 45 and Kimi K3 at 44, and level with xAI's closed Grok 4.7. It costs about $0.13 per index task, against $2.00 for Kimi K3, so it is both better and roughly fifteen times cheaper to use than the model it displaced.
Xiaomi's own tables put it close to the closed frontier models of mid-2026. It scores 71.9 on DeepSWE against 74.0 for Claude Opus 5, 89.9 on Terminal Bench 2.1 against 89.1, and 53.1 on AutomationBench, ahead of all three closed models it lists. On the harder Terminal Bench 4.0 it trails clearly, 34.9 against 49.0, and on exploit-writing tests it is well behind. These are Xiaomi's runs and should be read as such.
Flash is the practical choice for most teams. It stays within a few points of Pro on most of Xiaomi's agent tests while costing a third as much per token and running on far less hardware. Pro is for teams that want the top open model and can afford to host it.
The weak points are speed and serving. Artificial Analysis measured 56.8 tokens a second on Xiaomi's API, slower than most models in its class. Xiaomi's SGLang recipe uses 16-way tensor parallelism across two nodes. Its own model card also documents a tendency to repeat tool calls in agent loops, which the later MOPD checkpoints were trained to reduce.
Where it earns its place
A startup running high-volume coding and research agents
Cut model costs without dropping far below closed-model quality
Pro scores within a few points of Claude Opus 5 on Xiaomi's agent tests at $0.435/$0.87 per million tokens. A task with 500K input and 30K output tokens costs about $0.24 before caching, against about $1.95 on Kimi K3.
An enterprise with its own GPU cluster
Run a frontier-class open model entirely in-house
Serve the MOPD checkpoint with SGLang across two 8-GPU nodes, or vLLM at tensor-parallel 8 for smaller loads. The MIT licence puts no conditions on commercial use.
A team building multimodal agents
Handle screenshots, video and audio in one model
Both Pro and Flash take image, video and audio input natively. Flash, at $0.14/$0.28, is the cheaper starting point and scores 80.8 on OSWorld-Verified in Xiaomi's tests.
Before you commit
- Setup
- Call it through the Xiaomi MiMo API platform or OpenRouter, or download the weights from Hugging Face or ModelScope. Xiaomi recommends the MOPD checkpoints for agent use and temperature 1.0, top_p 0.95.
- Integrations
- SGLang and vLLM recipes with MiMo tool-call and reasoning parsers; MiMo Studio, MiMo Code and MiMo Desktop apps; OpenRouter.
- Data and privacy
- Weights under MIT. Xiaomi's MiMo privacy policy says data is stored in Singapore and deleted or anonymised after its retention period; the Desktop app trains on inputs unless you opt out. Third-party gateways report that Xiaomi's API does not train on traffic, which Xiaomi's API terms should confirm.
- Cost at scale
- $0.435/$0.87 per million tokens with a 99% cache discount and half-price batch. 1,000 agent tasks of 500K input and 30K output tokens cost about $244 uncached; Flash would be about $78.
- Limits
- 1M-token context. Artificial Analysis measured 56.8 tokens a second and a 3.11-second first token on Xiaomi's API. MiMo-V2.5 and V2.5-Pro leave Xiaomi's API on 21 October 2026.
What's good
- First among open models on Artificial Analysis's Intelligence Index (46)
- About $0.13 per index task, against $2.00 for Kimi K3
- MIT licence, with no revenue or display conditions
- 1M-token context and text, image, video and audio input
What's not
- Slow at 56.8 tokens a second on Xiaomi's API
- Self-hosting Pro needs two 8-GPU nodes
- Trails Claude Opus 5 on Terminal Bench 4.0 and exploit tests
- Original checkpoints repeat tool calls in agent loops
Best for
Teams that want the strongest open-weight model for agents and coding, either hosted cheaply on Xiaomi's API or self-hosted on a GPU cluster.
MiMo-V2.6-Pro alternatives
MiMo-V2.6-Pro: common questions
- Is MiMo-V2.6-Pro worth it?
- Xiaomi now makes the top-scoring open-weight model, at a fraction of closed-model prices; the catch is that self-hosting it is a 16-GPU job. We rate it 4.5 out of 5.
- How much does MiMo-V2.6-Pro cost?
- Open weights under MIT. Xiaomi API: Pro $0.435 input / $0.87 output per 1M tokens, cached input $0.004; Flash $0.14 / $0.28; Pro UltraSpeed $4.35 / $8.70; Batch API half price.
- What are the downsides of MiMo-V2.6-Pro?
- Slow at 56.8 tokens a second on Xiaomi's API; self-hosting Pro needs two 8-GPU nodes; trails Claude Opus 5 on Terminal Bench 4.0 and exploit tests; original checkpoints repeat tool calls in agent loops.
- Who is MiMo-V2.6-Pro best for?
- Teams that want the strongest open-weight model for agents and coding, either hosted cheaply on Xiaomi's API or self-hosted on a GPU cluster.
- How much does MiMo-V2.6-Pro cost for a team or in production?
- $0.435/$0.87 per million tokens with a 99% cache discount and half-price batch. 1,000 agent tasks of 500K input and 30K output tokens cost about $244 uncached; Flash would be about $78.
- Is MiMo-V2.6-Pro safe for company data?
- Weights under MIT. Xiaomi's MiMo privacy policy says data is stored in Singapore and deleted or anonymised after its retention period; the Desktop app trains on inputs unless you opt out. Third-party gateways report that Xiaomi's API does not train on traffic, which Xiaomi's API terms should confirm.
- What are the best MiMo-V2.6-Pro alternatives?
- The closest alternatives we have reviewed are GLM-5.3, Kimi K3.1 Pro, DeepSeek-V4-Pro and Qwen3.8-Max.
Answers checked Oct 25, 2026.
Make MiMo-V2.6-Pro? Show this rating on your site
