Xiaomi's MiMo-V2.6 notice describes only V2-Flash
A Xiaomi MiMo news page headlined as the release of three MiMo-V2.6 models contains no details on them, describing only MiMo-V2-Flash, the open-weight MoE model first released in December 2025.
A news page on Xiaomi's MiMo site is titled "Xiaomi MiMo-V2.6 Series: 3 New Models Officially Released," but its body doesn't name, describe or benchmark any V2.6 model. The page sits under a "previous news" path dated 20251216. Its text is the launch announcement for MiMo-V2-Flash, so as of this writing, the V2.6 release it advertises has not been confirmed.
The material that is there covers MiMo-V2-Flash, a Mixture-of-Experts model with 309B total and 15B active parameters. Xiaomi released the weights and inference code under the MIT license. The company says the model ranks in the top two among open-source models on several agent benchmarks. It also claims the model's coding beats every other open model and is comparable to Claude 4.5 Sonnet, at about 2.5% of Claude's inference cost and twice the generation speed. These are Xiaomi's own figures.
API pricing is listed at $0.10 per million input tokens and $0.30 per million output tokens. The API was offered free for a limited time.
On architecture, the model uses hybrid attention: one global-attention layer for every five sliding-window layers, with a 128-token window. That keeps the KV cache at a fixed size. The model was pre-trained on 32k-token contexts and extended to 256k. A three-layer multi-token prediction setup gives a reported 2.5×–3.7× real-world inference speedup.
The page links to a technical report, Hugging Face weights, a GitHub repository and an LMSYS blog post, all for V2-Flash. Xiaomi hasn't published names, sizes, benchmarks or availability for the three V2.6 models.