Yes, for English or Chinese work on a budget. MiMo-V2.6-Pro is the highest-scoring open-weight model on the Artificial Analysis index, and it scores 55.20% on the Vals Index at a cost per task that undercuts almost everything else. Its limits are language coverage, since Xiaomi’s cards list only English and Chinese, and a very short track record: it came out on 22 September 2026.
First principles
Four things decide whether this model is worth testing: what a task costs, what it takes to host, which languages it learned, and how much evidence there is so far.
Cost per task, not price per token. A task costs the tokens the model spends multiplied by the price of each token, so a model that finishes in fewer tokens can beat one with a lower rate card. That is why this review relies on the cost-per-task figures from Artificial Analysis and Vals. It matters more here because Xiaomi’s release notes give no prices.
Active parameters set the work per token; total parameters set the memory. MiMo-V2.6-Pro is a mixture-of-experts model: each token runs through 42B of its 1.02T parameters, but all 1.02T must be loaded. Even at 8 bits per parameter that is about one terabyte of weights, which means a server with several graphics processing units (GPUs), not a workstation.
Languages come from training data. A model is only as good in a language as the text it learned from. Xiaomi lists English and Chinese, and a high score on general benchmarks does not show quality in a language the vendor does not list.
A week is a small sample. Leaderboard positions for a new model can move as more votes and runs come in, and a week of public use says little about how it holds up in production. Treat the rankings as provisional and check them again in a month.
What it is
MiMo-V2.6 is Xiaomi’s model series, released on 22 September 2026 in three hosted tiers: mimo-v2.6-pro, mimo-v2.6-flash and a faster mimo-v2.6-pro-ultraspeed. Xiaomi also published open weights under the Massachusetts Institute of Technology (MIT) licence.
- MiMo-V2.6-Pro is a mixture-of-experts model with 1.02T total and 42B active parameters and a 1M token context window. It reads text, images, video and audio in one model and returns text.
- MiMo-V2.6-Flash has 309B total and 15B active parameters, with the same input types and window.
- MiMo-V2.6-Distill-Qwen-9B is a small model fine-tuned from Qwen3.5-9B, also MIT, which reads text and images.
Xiaomi says it trained the Pro in a single reinforcement learning run across coding, agents, vision and cybersecurity, rather than separate runs per domain.
Where it is good
Independent rankings:
- Artificial Analysis gives MiMo-V2.6-Pro 46 on its Intelligence Index, first on its open-weights page, ahead of GLM-5.3 (45) and Kimi K3 (44). Its cost to run one index task was $0.13 on the leaderboard page. The open-weights page on the same site shows $0.2, so treat the figure as approximate.
- The Vals Index (updated 29 September 2026) gives it 55.20% at $0.41 per test. That is above Grok 4.7 (54.95% at $12.12) and Gemini 3.8 Flash (54.83% at $5.73). The Flash tier scored 53.23% at $0.20.
- On the Arena text board (updated 25 September 2026) it ranks 23rd at 1480.
Cost per task is the standout. On Vals, apart from its own Flash tier, only DeepSeek-V4.1-Flash and GPT-6 Luna come near its price at similar scores, and MiMo-V2.6-Pro scores higher than both.
Input range. Few open models take audio as well as images and video. Xiaomi reports 89.9 on Terminal Bench 2.1 and 71.9 on DeepSWE v1.1 for the Pro, both vendor-reported.
Small option. The 9B distilled model gives teams a much smaller MiMo-style model to host. Xiaomi reports it lifts AutomationBench from 5.0% to 30.3% over its Qwen base, a vendor-reported figure.
Where it is bad
Languages. The Flash model card lists English and Chinese as supported languages. We found no broader list for the Pro. On this evidence it is a poor choice for Hindi or any other Indian language.
Speed. Artificial Analysis measured about 41 output tokens per second. On Vals, an average run took 1 hour 7 minutes, against about 28 minutes for GPT-6 Astra.
Short record. One week of public use is not enough to know how it holds up in production. Independent coverage is limited to the boards above.
Pricing is unclear. Xiaomi’s release notes do not give prices, and we could not confirm a hosted rate, so we give none.
Size. The Pro’s 1.02T parameters need a server with several GPUs to host. The Flash, at 309B, is lighter but still well beyond a single consumer GPU.
Price and access
Weights for Pro, Flash and the 9B are on Hugging Face and ModelScope under MIT. The Flash card lists SGLang, vLLM, Ollama and llama.cpp support. Xiaomi offers hosted access through its own application programming interface (API) platform, AI Studio and desktop app.
| Model | Total / active parameters | Inputs | Licence |
|---|---|---|---|
| MiMo-V2.6-Pro | 1.02T / 42B | Text, image, video, audio | MIT |
| MiMo-V2.6-Flash | 309B / 15B | Text, image, video, audio | MIT |
| MiMo-V2.6-Distill-Qwen-9B | 9B class | Text, image | MIT |
Who should use it
On this evidence MiMo-V2.6-Pro is the open model to test first for English or Chinese agent work where cost matters. Confirm prices with Xiaomi before you plan around it, and re-check results in a month. For Indian languages, look at Sarvam instead.
Sources
- Xiaomi MiMo-V2.6 Series: 3 New Models Officially Released, read 2026-09-30
- MiMo-V2.6 collection on Hugging Face, read 2026-09-30
- MiMo-V2.6-Pro-RL on Hugging Face, read 2026-09-30
- MiMo-V2.6-Flash-RL on Hugging Face, read 2026-09-30
- MiMo-V2.6-Distill-Qwen-9B on Hugging Face, read 2026-09-30
- LLM Leaderboard (Artificial Analysis), read 2026-09-30
- Top Open Source AI Models (Artificial Analysis), read 2026-09-30
- Vals Index, read 2026-09-30
- Arena text leaderboard, read 2026-09-30