MiMo-V2.6 open weights: frontier scores, MIT license, 1M context

Xiaomi's MoE pair tops the open-weights charts: Flash runs near-Pro at a third the cost, and the drop bundles 7,000+ RL environments and a full training stack.

Nowline SEP 25 8:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • MiMo-V2.6-Pro: the open model that beats DeepSeek

    A 1.02T-param sparse MoE (42B active) with a 1M-token context and native text, image, audio and video input. It posts DeepSWE 71.9 and Terminal Bench 89.9 — frontier-class scores, MIT-licensed, and downloadable from Hugging Face today, so you can self-host or fine-tune it with no usage restrictions.

  • MiMo-V2.6-Flash: the one you'll actually deploy

    310B total / 15B active, so it fits far smaller hardware while nearly matching Pro (DeepSWE 67.9, Terminal Bench 87.6). Via API it runs about $0.14 in / $0.28 out per million tokens — roughly a third of Pro's price and cheap enough to put behind a real product.

  • It ships with 7,000+ RL environments and a training stack

    This isn't just weights. Xiaomi bundled 7,000+ reinforcement-learning environments and an end-to-end agent-training framework, so you can extend the model's tool-use and coding skills yourself — spin up a specialized agent this weekend instead of only prompting the base model. Serving recipes for vLLM are already published.

  • Elsewhere: Grok 4.7 arrives with a bigger base, same price

    xAI shipped Grok 4.7 — a larger base model trained with extended RL, holding Grok 4.6's $2/$6 per-million pricing and 500K-token context. If you're already on the Grok API, it's a drop-in bump with no cost change.