MiMo-V2.6-Pro

A 1T-parameter omni model whose training receipts are public. MiMo-V2.6-Pro is Xiaomi’s flagship open-weight model: 1.02T total parameters with 42B active per token (384 routed experts, 8 active), one model handling text, image, video, and audio with a 1M-token context window. The unusual thing is not the size, it is the disclosure: Xiaomi livestreamed the reinforcement learning run and published the bill. Flash cost about $850K, Pro about $2.62M, together $3.47M, under six days, 30 RL steps each, roughly 750,000 training trajectories. Frontier labs treat training cost as a trade secret; Xiaomi turned it into the marketing.

Why the RL run matters. The whole release is a bet that scaling RL compute (batches of 1,568 samples with 16 rollouts each, 3.5-3.7B tokens per step, a mixed task suite spanning code, general agents, visual work, and cybersecurity in one run) buys more than pretraining scale does. DeepSWE v1.1 jumped 14 points (58.4 to 72.6) from this single run. On the Artificial Analysis Intelligence Index the model scores 46, past Kimi K3 and Qwen3.8 Max, making it the top open-weights model at launch; Xiaomi itself says Claude Fable 5.1 and GPT-6 Astra (closed) still lead overall.

What runs where. Not your desk. The MoE memory rule is total parameters resident: 1.02T at 4 bits is ~508 GB before KV cache, and the vendor’s own serving recipe is SGLang across two nodes with 16-way tensor parallelism. You reach it through the API (priced same as the V2.5 series; Xiaomi claims 1/20 to 1/60 of overseas models at the same intelligence level), OpenRouter, or the MiMo Desktop app. Weights are MIT on HuggingFace (42,062 downloads in the first four days) for people with cluster budgets. The small sibling MiMo-V2.6-Flash (309B total) and a 9B Qwen distill exist; the 9B is the only member that approaches home hardware.

Benchmarks, labeled. Everything above is vendor-reported from the launch. The interesting rows for agent builders: AutomationBench v1.0.6 53.1 (vs Opus 5 at 50.3), Terminal Bench 2.1 89.9, CyberGym 94.0. Independent replication is still thin; the Artificial Analysis index is the only third-party signal at launch.

Why it matters beyond the scorecard. Two signals in one release: an open model beating every other open model, published for a training price that two people with a credit line could theoretically match, and a livestream that made the process inspectable. If the $3.47M receipt holds up as the real cost of a frontier-adjacent model, the “only megacaps can play” premise weakens another notch.