Xiaomi has released a new lineup of large language models called MiMo-V2.6. Led by a flagship Pro variant with over a trillion parameters, the family spans several sizes, and Xiaomi has made all of the weights freely available under the permissive MIT license. Independent benchmarking has placed the flagship at the top of the open-weight model rankings, drawing attention as an example of strong performance achieved on a comparatively modest development budget.
At the center of the lineup is MiMo-V2.6-Pro, which exceeds a trillion total parameters while relying on a mixture-of-experts design that only activates a fraction of that capacity for any given token. A smaller MiMo-V2.6-Flash variant, with total parameters in the hundreds of billions, trades some capability for a lighter computational footprint.
Xiaomi also released a hosted, latency-optimized version of Pro called Pro-UltraSpeed, which delivers output at up to 20 times the speed, alongside a sub-10-billion-parameter distilled model intended for research use. Pro, Flash, and Pro-UltraSpeed all share a one-million-token context window and can accept text, images, audio, and video as input, though output remains text-only for now.
A Mixture-of-Experts Core With Vision Support
Internally, Pro is built from dozens of transformer blocks routing through several hundred experts, only a handful of which fire for any single token. A speculative decoding mechanism helps speed up generation, and a dedicated vision encoder handles image understanding.
Its attention mechanism combines narrow, sliding-window attention with a smaller number of global attention layers that can see the entire sequence. For deployment, Xiaomi indicates that running Pro at full precision calls for roughly eight high-end GPUs, while Flash can run on about half that number, giving companies a rough benchmark for self-hosting.
Training Costs Made Public, With Efficiency Front and Center
One of the standout claims around MiMo-V2.6 is how little it reportedly cost to train. Xiaomi says it ran a single large-scale reinforcement learning pass covering coding, general agentic tasks, visual reasoning, and cybersecurity all at once, at a combined cost of around 3.5 million dollars (roughly 553 million yen). The training process relied on an efficient reinforcement learning algorithm that samples large batches of prompts in parallel and asynchronously, squeezing more value out of the available compute. Xiaomi also took the unusual step of livestreaming the training run itself.
※1 USD = 158 JPY (as of September 25, 2026)
On a widely referenced third-party intelligence index, Pro reportedly earned the top score among open-weight models. Its per-task inference cost also came in under a dollar, underscoring the balance Xiaomi is striking between capability and cost.
Benchmarks Approach the Frontier
Individual technical benchmarks tell a similar story. Across evaluations covering software engineering, agentic task completion, and command-line operation, Pro consistently outperformed the smaller Flash model, with both showing practical, real-world levels of competence. Reported scores on more complex, file-manipulation-heavy tasks and security-oriented evaluations were similarly solid.
That said, some reports note that Pro still trails the very best proprietary models by a few points on certain benchmarks. Even so, given that its weights are freely available and its training budget was comparatively small, MiMo-V2.6 carves out a distinct niche by pairing strong performance with openness.
Broad Availability, Pay-as-You-Go Pricing
MiMo-V2.6's weights are distributed on Hugging Face under the MIT license, permitting both research and commercial use. Beyond that, the models are also accessible through Xiaomi's own API as well as several third-party AI platforms, giving developers flexibility in how they integrate the models into their own stacks.
Pricing through third-party platforms puts Pro's per-million-token input cost under a dollar, with output priced proportionally. The lighter Flash model is priced even lower, while the latency-optimized Pro-UltraSpeed carries a premium in exchange for its speed. Some developer tools also ran limited-time promotions offering free access to the Flash model shortly after launch.
Summary
Xiaomi's newly announced MiMo-V2.6 family stands out chiefly for combining a trillion-parameter-class model, released freely under the MIT license, with a training budget of roughly 3.5 million dollars. Independent evaluations have placed it at the top of the open-weight rankings, and it is available through direct weight downloads as well as multiple API options. While it still falls a little short of the very top proprietary models on some measures, its combination of low development cost and competitive pricing positions it as a compelling option in the increasingly crowded field of open AI models.
Sources
https://datanorth.ai/news/xiaomi-releases-mimo-v2-6-pro-and-flash
