Nearly half a year of silence. We spent it studying one problem: how far RL can scale. MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks. Streaming the run:

Recently I sat down with Matt White, former Global AI CTO at the Linux Foundation and former CTO of the PyTorch Foundation. We got into open-source AI, model licensing, agent infrastructure, and what it takes to build a lasting open-source community. Really enjoyed this one. Thanks again, Matt! @matthew_d_white Watch here:

Hosting something open-source during OSAI Week? Meetup, workshop, demo — get it on the community calendar. Submissions open now. #OpenSourceAIWeek# #OSAIWeekCommunity#

Hosting something open-source during OSAI Week? Meetup, workshop, demo — get it on the community calendar. Submissions open now. #OpenSourceAIWeek# #OSAIWeekCommunity#

Anthropic has lined up compute deals that could cost as much as $517 billion, far above the $180 billion it told investors it expected to spend through 2029. The commitments underscore how aggressively it is preparing for growth ahead of a possible IPO. More insights:

Can open source prevent AI power from concentrating? MTS host @sophiadew visits the Open Source AI Summit in San Francisco to ask researchers and founders working at different layers of AI. Co-author of Attention Is All You Need, @lukaszkaiser, argues power concentration is a property of the transformer, not a permanent feature of AI. Former Linux Foundation Global CTO of AI, @matthew_d_white unpacks why China leads on open weights and what the US needs to compete. @VictorSuOrtiz shares what open weights unlock for creators building on @MiniMax_AI's H3 video model. Cofounder and CTO of @FireworksAI_HQ, @dzhulgakov breaks down how Fireworks AI allows companies to own their intelligence And NVIDIA's @JeanKossaifi argues that nearly every major breakthrough in modern AI already rests on an open foundation. 1:20 Lukasz on why AI power concentrated 4:12 One RTX 5090 > 8-GPU setup his team used for transformer research 8:02 Matt on why US and China have opposite incentives on open source 11:44 Scarcity forces China to innovate, abundance lets the US scale up 15:43 Why Poolside stopped pre-training 20:18 Victor on the H3 community shipped 21:49 Dima on renting intelligence vs training your own 27:01 How training on your own data gets better results from a smaller, cheaper model 30:40 Jean Kossaifi on tracing modern AI back to the Python ecosystem