MiMo-V2.6-Flash

Overview. MiMo-V2.6-Flash is Xiaomi's efficiency-tier reasoning model from the MiMo-V2.6 series, released in September 2026 under an MIT licence with published weights. It is a sparse Mixture-of-Experts model with 309B total parameters and 15B activated per token, drawing on 256 routed experts of which 8 fire per token, and it understands text, images, video and audio natively while returning text. A developer reaches for it when they need long-context reasoning over mixed media without stepping up to the flagship tier of the family.

Key strengths. The architecture balances capacity against serving cost: 309B total parameters behind only 15B active per token, plus a 1M-token context window with up to 128K tokens of output. It is an RL-trained reasoning variant with a caller-controlled thinking switch, and its strongest evidenced area is long-context, multimodal research work, where the curated domain signal leads the rest of the vector. On the Artificial Analysis Intelligence Index it scores 37.9, a mid-tier placement consistent with its position below the family flagship rather than a claim of frontier quality.

When to use it. Use this model for reasoning tasks that also need wide context — synthesising and analysing long documents, images, video or audio in a single pass, where a 1M-token window and native streams of image and audio input remove the need for pre-processing pipelines. It fits interactive applications, with a throughput-oriented sparse MoE and a binary thinking switch that lets you confirm the reasoning trace on when a task needs it and off when it does not. Prompt caching is exposed, which helps when the same long context is reused across calls.

Limitations. Tool calling and structured output are documented for this model, but parallel tool-call behaviour is not clearly specified, so multi-step agentic use should be treated as provisional. Reasoning is a binary on/off switch with no selectable effort levels, and while thinking is enabled the vendor fixes temperature and top-p, so sampling control is limited in that mode. Output is text only — the model understands audio and video but does not generate either — and Xiaomi has not published a training cutoff.

Sovereignty note. The creator is Xiaomi, based in China, and the weights are published under an MIT licence, which means the model can be self-hosted on infrastructure you control if you need full inference sovereignty. The Stav-hosted endpoint is delivered by DeepInfra, a US-incorporated provider, so that hosted endpoint does carry US CLOUD Act exposure; buyers in regulated EU jurisdictions should still confirm procurement and jurisdiction fit.