Can LLMs adapt continually without losing base skills? Fast-Slow Training (FST) pairs "slow" weights with "fast" context. FST vs. RL: • 3x more sample-efficient • Higher performance ceiling • Less KL drift (better plasticity) • Continual learning: succeeds where RL stalls

Super excited to announce seven new world-class MAI models today. They represent what we consider a new era in AI designed to keep you in control and on the frontier. First is our text foundation model, MAI-Thinking-1, exceptionally strong on reasoning and SWE tasks. - It’s a 35B active parameter MoE with a 256K context window. Independent human raters on Surge prefer it for overall quality in blind side-by-sides versus Sonnet 4.6, and it’s achieved 97% on AIME 2025, the key measure of its general-purpose reasoning abilities. - It's at 53% on SWE Bench Pro, placing it right alongside Opus 4.6 on one of the toughest coding benchmarks. - And since we co-designed our models with our own silicon, MAI-Thinking-1 is optimized on our MAIA 200 chip. Benchmarking head-to-head against the GB200, we see 30% better performance per dollar as well as a 1.4x performance-per-watt gain when running our MAI models on the MAIA 200 end-to-end. Next is MAI-Image-2.5 and its Flash variant. Two super strong models now at #2 on the leaderboards, surpassing the score of Nano Banana 2 on image editing. Last for now is MAI-Code-1-Flash, our new inference efficient coding model, especially tuned for VS Code and GitHub Copilot CLI. - Code-1-Flash achieves 51% on SWE Bench Pro, despite having just 5B parameters, putting it closer to Haiku in size but cheaper in cost. All of this is the foundation for Microsoft Frontier Tuning. It lets you customize our models to create custom, company-specific agents that only you control. You can make our model, your model. Your data. Your agents. Your moat. Early adopters are already seeing a difference. When we tuned our models for McKinsey’s tasks, MAI delivered the highest win rate, outperforming GPT-5.5 on quality, while being 10x lower on cost. Also really excited to be collaborating with the amazing team at Mayo Clinic to jointly train a new frontier AI model for healthcare. Our announcements today mark another milestone on the road to humanist superintelligence. You can learn more and about our other new models in our latest blog: microsoft.ai/news/building-a…

Excited that our paper StreamdiffusionV2 received the Best Research Paper Award at #MLSys26! 🚀Video generation is quickly moving from demos to production-facing workloads. It is no longer a turn-based pipeline but should be a streaming pipeline to interact with users. 📖Our project page: streamdiffusionv2.github.io/ and paper: arxiv.org/pdf/2511.07399 👂Come join the talk if you are interested in streaming video generation. Our talk will be at the Research Track Oral Presentation: Best Paper Session on Tue 8:45AM at #MLSys26 , I will talk about how we attacked the efficiency and quality challenges. Hope to see you there! ❤️Huge thanks to all authors! This work would not have been possible without the incredible effort from the entire team. Big shout out to Tianrui Feng, Zhi Li, @Andy_ShuoYang , @HaochengXiUCB, @lmxyy1999 , @lvminzhang , @xiuyu_l , Keting Yang, @ZiqiPeng, @songhan_mit , @magrawala, @KurtKeutzer , and @cumulo_autumn

Learning from rich textual feedback (errors, traces, partial reasoning) beats scalar reward alone for LLM optimization. GEPA demonstrated this for context-space optimization (prompts and agent harnesses), delivering frontier results at a fraction of the cost of RL. But context-only optimization is bounded by the base model's capability ceiling; weight updates can reach further. Very excited about this new line of work on Fast-Slow Training (FST), which interleaves context and model weight optimization! The idea is a clean division of labor between two interleaved loops: 🔹 Fast loop (context): GEPA reads rich rollout feedback updating the context layer. The context becomes a fast-updating scratchpad of what the model needs to know about this task, right now. 🔹 Slow loop (model parameters): RL updates the model's parameters conditioned on the evolving context. Because the prompt already carries task-specific nuances, the model parameters are freed from absorbing them and focus on what actually generalizes across tasks and pushes the frontier. ⦁ 3× more sample-efficient than RL on math, code, and physics reasoning ⦁ ~70% lower KL divergence from base at matched accuracy ⦁ Plasticity preserved: FST checkpoints respond better to additional RL on new tasks than RL-only ones ⦁ Continual learning across changing tasks (HoVer → CodeIO → Physics) where RL stalls the moment the task switches FST is a direction towards: ⦁ Addressing RL's pain points: entropy collapse, sparse rewards, long-horizon exploration ⦁ Providing a clean channel for rich feedback into weight updates ⦁ Demonstrating model-harness co-evolution ⦁ Discovery: Using fast context updates for broad exploration, while leveraging a continually improving model. Check out the full thread below:

Can LLMs adapt continually without losing base skills? Fast-Slow Training (FST) pairs "slow" weights with "fast" context. FST vs. RL: • 3x more sample-efficient • Higher performance ceiling • Less KL drift (better plasticity) • Continual learning: succeeds where RL stalls

Training LLMs is synonymous with updating their weights. However, LLMs can also learn in-context using *frozen* weights. There is no good reason for restricting learning to being in-context or in-weights. So a natural idea is "Learning, Fast and Slow" (FST). In FST, slow learning is LLM weights trained with RL while fast learning is context / prompt (fast weights) optimized with GEPA. Compared to RL, FST performs better while being more data efficient, adaptable (plasticity), and forgetting less (stays closer to base models). I think this idea of learning both fast-slow weights would be a good foundation for continual learning. PS: Geoff Hinton (the OG) described the idea of fast weights and slow weights several years ago, and back then I remember thinking it's a very cool idea. See more details here: gepa-ai.github.io/gepa/blog/…

Very excited about this line of research of fast-slow learning, 1) potential to solve a lot of issues with current RL (eg. entropy collapse, sparse rewards) 2) an intuitive way of incorporating rich feedback with RL 3) provides a way to transfer knowledge of text-only based learning into the model 4) a great candidate for model-harness co-evolution, seeing a lot discussion on X lately about future models developing their own harness. 5) most importantly, can imagine these kinds of algorithms to be more suitable candidates for discovery that requires both extreme exploration but at the same time improving the underlying model capabilities. and much more ...

Training LLMs is synonymous with updating their weights. However, LLMs can also learn in-context using *frozen* weights. There is no good reason for restricting learning to being in-context or in-weights. So a natural idea is "Learning, Fast and Slow" (FST). In FST, slow learning is LLM weights trained with RL while fast learning is context / prompt (fast weights) optimized with GEPA. Compared to RL, FST performs better while being more data efficient, adaptable (plasticity), and forgetting less (stays closer to base models). I think this idea of learning both fast-slow weights would be a good foundation for continual learning. PS: Geoff Hinton (the OG) described the idea of fast weights and slow weights several years ago, and back then I remember thinking it's a very cool idea. See more details here: gepa-ai.github.io/gepa/blog/…

Can LLMs adapt continually without losing base skills? Fast-Slow Training (FST) pairs "slow" weights with "fast" context. FST vs. RL: • 3x more sample-efficient • Higher performance ceiling • Less KL drift (better plasticity) • Continual learning: succeeds where RL stalls

Can LLMs adapt continually without losing base skills? Fast-Slow Training (FST) pairs "slow" weights with "fast" context. FST vs. RL: • 3x more sample-efficient • Higher performance ceiling • Less KL drift (better plasticity) • Continual learning: succeeds where RL stalls

Can LLMs adapt continually without losing base skills? Fast-Slow Training (FST) pairs "slow" weights with "fast" context. FST vs. RL: • 3x more sample-efficient • Higher performance ceiling • Less KL drift (better plasticity) • Continual learning: succeeds where RL stalls