Context
Xiaomi has released and open-sourced the MiMo-V2.6 series: two natively omni-modal models —MiMo-V2.6-Pro and MiMo-V2.6-Flash— plus an UltraSpeed variant that delivers up to 20x faster output at the same quality. With 46.32 on the Artificial Analysis Intelligence Index, MiMo-V2.6-Pro ranks as the strongest open-source model to date, ahead of Kimi K3 and Qwen3.8 Max, while keeping the API pricing of the previous generation.
The headline for us isn't the ranking itself: it's that the intelligence-per-cost frontier moved outward again. In a multi-model system, that means harder tasks solved at a lower cost per task.
What changes
- Strong on long-horizon agentic work: 71.9 on DeepSWE v1.1 (software engineering), 76.9 on Toolathlon-verified and 62.0 on JobBench. In security domains such as CyberGym, the series actually beats several closed reference models.
- Omni-modal by design: text, image, video and 3D spatial reasoning live in the same model — no separate pipelines, no middleware connectors.
- Three profiles to route across: Pro for the hard calls, Flash for volume, and an UltraSpeed mode for when latency is the constraint.
- Open training: Xiaomi published the technical report, the training environments and the RL code, and streamed the production run live (~750k trajectories, 30 RL steps, with gains that kept compounding throughout).
The charts that matter
From "vibe coding" to "vibe world"
MiMo's coding ability generalized well beyond software: given a goal, a harness and time, it builds, tests and ships. Combined with multimodal perception and computer-use agent (CUA) capabilities, the series moves from writing software to building interactive worlds — 3D scenes, interaction logic and visual verification, iterating on the rendered output.
[
](https://mimocode-cdn.xiaomimimo.com/mimocode/blog/mimo-v2-6/02.mp
4)
[
](https://mimocode-cdn.xiaomimimo.com/mimocode/blog/mimo-v2-6/04.mp
4)
[
](https://mimocode-cdn.xiaomimimo.com/mimocode/blog/mimo-v2-6/07.mp
4)
What it means for you
If you already run on NapsixB2B, nothing in your operation changes: the MiMo-V2.6 series joins the model pool and the router picks it when it's the right fit. Concretely:
- More ceiling, no bigger bill: coding, analysis and automation work that used to require a heavyweight model can now run on MiMo-V2.6-Pro at the same API pricing as the previous generation.
- Cheap volume: MiMo-V2.6-Flash covers repetitive work — classification, extraction, triage, summaries — at near-frontier quality and a fraction of the cost.
- Speed when it matters: UltraSpeed targets interactive flows where generation speed is part of the experience.
- Genuine multimodality: documents with charts, screenshots, video or spatial references no longer need separate preprocessing.
What we do
We benchmark every new model against our own evaluations before enabling it, and only add it to the router when it improves the intelligence–cost trade-off for some family of tasks. MiMo-V2.6 cleared that bar: it enters as a first-line option for coding, general agents and visual tasks, under the same policy as always —efficiency, not brute force: every task goes to the right-sized model, and the free-model floor stays always on, 24/7, multi-agent and multi-user.
