Xiaomi is better known for phones and rice cookers than for foundation models, but it just released two of them anyway. MiMo-V2.6 Pro and MiMo-V2.6 Flash are omnimodal systems built for coding, visual reasoning, and controlling a computer directly, and Xiaomi has open-sourced both. A third variant, Pro-UltraSpeed, promises the same output quality at up to 20 times the speed for latency-sensitive tasks.
On the Artificial Analysis Intelligence Index v4.3, Xiaomi says Pro scored 46.32, ranking above Moonshot’s Kimi K3 and Alibaba’s Qwen3.8 Max as the top open-weight score in that comparison. That ranking comes from Xiaomi’s own release materials, not an independent audit. In a launch post on X, the MiMo team also said Pro “performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks.” Anthropic and OpenAI have not confirmed that comparison, and it should be read as a vendor’s framing until a third party reproduces it.
The pricing is where Xiaomi is making its real argument. Uncached input costs $0.14 per million tokens on Flash and $0.435 on Pro; output runs $0.28 and $0.87 respectively. Both rates match the prior V2.5 generation, and UltraSpeed costs ten times more for the speed premium. For a developer choosing between a frontier-labeled open model and a closed one from Anthropic or OpenAI, that is a meaningfully cheaper entry point, assuming the benchmark claims hold up under outside testing.
The training process is the more unusual disclosure. Coding, general-agent work, vision, and cybersecurity tasks all fed the same push: reinforcement learning scaled against problems whose outcomes can actually be verified, using context windows as long as one million tokens. Each model ran 30 RL steps over roughly 750,000 trajectories in under six days: Flash cost about $850,000 to train that way, Pro about $2.62 million. On DeepSWE v1.1, Xiaomi reports Flash’s score rising from 48.8 to 65.68 and Pro’s from 58.4 to 72.57 over that run. Reward hacking, the failure mode where a model learns to game its own scoring function rather than solve the task, was the risk Xiaomi says it built against: the routing layer was frozen mid-training to limit drift, and verifier cross-checks, anomaly detection, and adversarial evaluation ran alongside it.
Xiaomi is pitching the models’ range past ordinary coding work under the label “Vibe World”: from a single image, video, or text prompt, the system can direct agents to build and visually test 3D scenes, generate Blender assets, drive a Franka Panda robotic arm from live camera input, draft frontends and slide decks, cut video, and write music as both a score and a MIDI file. Xiaomi also points to research demos in chemistry (screening materials meant to capture PFAS compounds) and formal mathematics (a Lean 4 theorem verified across more than 6,000 lines of kernel-checked code), though these read as showcase runs rather than benchmarked results.
Both models are live now on AI Studio, MiMo Code, MiMo Desktop, Xiaomi’s own API platform, and OpenRouter, with MiMo Desktop exiting early access alongside the release. Xiaomi is also publishing the technical report, the training environments, and the RL code itself, which lets outside teams check the training-cost and benchmark numbers rather than take them on faith.
If Pro’s benchmark position survives independent testing, it adds a fourth serious open-weight contender, alongside DeepSeek, Qwen, and Kimi, to a segment now competing on RL training efficiency rather than raw parameter count. Teams evaluating open models on cost should treat Xiaomi’s per-token pricing as the number worth re-checking first, since it undercuts the closed frontier labs by a wide margin if the quality claim holds.
Reported by TestingCatalog on September 21, 2026.
