Xiaomi releases MiMo-V2.6 omni-modal models with MIT-licensed weights and disclosed RL training costs
The AMW Read
Extends the two-days-prior top-open-weight claim with rare RL-cost transparency and reproducible training code, meaningfully updating the open-weight competitive baseline against Claude Opus 5/GPT-5.6 without resolving a debate or introducing a new player.
Xiaomi releases MiMo-V2.6 omni-modal models with MIT-licensed weights and disclosed RL training costs
Xiaomi released its MiMo-V2.6 series on September 22, comprising the flagship MiMo-V2.6-Pro and the cost-efficient MiMo-V2.6-Flash, both handling text, image, video, and audio in a single omni-modal model. Pro carries 1.02 trillion total parameters with 42 billion active per token in a sparse mixture-of-experts design and a 1-million-token context window; a Pro-UltraSpeed variant delivers up to 20x faster output at matching quality. Weights for both models are released under an MIT license via Hugging Face. API pricing holds flat versus the V2.5 series: Flash runs $0.0028 per million input tokens on a cache hit ($0.14 on a miss, $0.28 output), Pro runs $0.0036/$0.435 input ($0.87 output). Both are live day-one across AI Studio, MiMo Code, MiMo Desktop (exiting early access with this release), the MiMo API Platform, and OpenRouter.
Xiaomi paired the release with unusual transparency: Flash and Pro each finished 30 reinforcement-learning steps on roughly 750,000 trajectories in under six days, at a disclosed cost of about $850,000 and $2.62 million respectively, with the technical report, training environment, and RL code published for outside reproduction. On Artificial Analysis's Intelligence Index, Pro scored 46.32 — the highest logged for an open-weight model, ahead of Kimi K3 and Qwen3.8 Max — and it edged Claude Opus 5 and GPT-5.6 Sol on the AutomationBench workflow benchmark (53.1 vs. 50.3 and 45.8), though it trailed both on the DeepSWE coding benchmark and on Terminal Bench. Per the AI Market Watch index, Xiaomi is tracked primarily under embodied AI/robotics, with name-matched pipeline mentions holding roughly steady (6 items in the last 90 days versus 8 in the prior 90, coverage limited to ingested sources) even as its foundation-model push accelerates.
For builders, MIT-licensed weights, flat pricing, and published RL code make Pro and Flash a cheap base for agentic and automation workloads, but the coding-benchmark gap against Opus 5 and GPT-5.6 argues for task-specific benchmarking rather than assuming parity across the board. For investors, a $2.62 million disclosed RL bill buying frontier-adjacent scores on select benchmarks reinforces how compressed the cost of catching up on evaluation leaderboards has become, shifting more of the durable advantage toward distribution and workflow integration rather than raw model quality.
#Xiaomi #MiMoV26 #OpenWeightModels #FoundationModels #ReinforcementLearning #AIBenchmarks
