1/n 🚀 Introducing SkyRL-Agent, a framework for efficient RL agent training. ⚡ 1.55× faster async rollout dispatch 🛠 Lightweight tool + task integration 🔄 Backend-agnostic (SkyRL-train / VeRL / Tinker) 🏆 Used to train SA-SWE-32B, improving Qwen3-32B from 24.4% → 39.4%

TurboDiffusion: 100–205× faster video generation on a single RTX 5090 🚀 Only takes 1.8s to generate a high-quality 5-second video. The key to both high speed and high quality? 😍SageAttention + Sparse-Linear Attention (SLA) + rCM Github: https://t.co/ybbNBjgHFP Technical

🤖 I am in San Diego for #NeurIPS2025 this week! Excited to chat about SkyRL(-Agent), Coding LLM/Agent, Self-evolving Agent, RL, and Inference/Training Infrastructure.

@BerkeleySky @anyscalecompute @awscloud @LambdaAPI @thinkymachines @DachengLi177 @fangz_zzu @sumanthrh @connorzchen @charlie_ruan @tyler_griggs_ @shulynnliu @erictang000 @CyrusHakha @richliaw @pcmoritz @matei_zaharia @profjoeyg @istoica 9/n — One more thank-you tweet because y’all earned it. Thanks to @jyangballin @StringChaos @jiayi_pirate @xingyaow_ for valuable and helpful feedback and discussions🔥

Check out our efficient infra for agentic RL training! More applications coming soon!🔥

Finally out! 😁

@BerkeleySky @anyscalecompute @awscloud @LambdaAPI @thinkymachines 8/n — Continued thanks. We also want to recognize the incredible team behind this work: @shiyi_c98 @DachengLi177 @fangz_zzu Shuo Yuan @sumanthrh @connorzchen @charlie_ruan @tyler_griggs_ @shulynnliu @erictang000 @CyrusHakha @richliaw @pcmoritz @matei_zaharia @profjoeyg @istoica

Proud to have contributed to SkyRL-Agent as an undergrad! Huge thanks to @shiyi_c98 and @DachengLi177 for all the guidance, learned a lot from this project. More details in the thread

7/n — Acknowledgements This work is developed in @BerkeleySky. In addition to the authors, we would like to thank all related open-source projects, and generous compute support from @anyscalecompute @awscloud @LambdaAPI @thinkymachines.

6/n — Join the Efforts & Roadmap SkyRL-Agent is a framework for efficient agent training. Looking forward, we are building: 📌 multi-agent training 📌 multi-domain training 📌 self-improving agents with runtime evolution It’s still an early stage, please join us to build

5/n — Other Case Studies SkyRL-Agent is not just for SWE. We also provide training examples for: 🧠 Deep Research Agent (document reasoning & evidence retrieval) 🖥 Computer Use Agent (OS operations) 📝 Memory Agent (memory management for long-context tasks) More recipes are

4/n Using SkyRL-Agent, we trained SA-SWE-32B purely with RL from Qwen3-32B. Training recipe highlights: 🔹 Trained with an AST-based search tool for better code navigation → ⚡ Higher Pass@K & sample efficiency due to improved tooling 🔹 Trained on 4.5K R2E-Gym instances,

3/n For SWE agent training, we use the Async Pipeline Dispatching method, which improves rollout throughput by 1.55× over naive async batching. Instead of leaving the GPU idle during CPU-bound stages (init, reward compute, etc.), pipeline execution better overlaps CPU + GPU

2/n SkyRL-Agent is built around three key components: 🧩 Tool-centric task interface  Supports dynamic registration of stateless tools, environment-modifying actions, and agent-state-modifying operations under a unified abstraction. ⚡ Efficient rollout scheduling  Fine-grained

Amazing project! Want to see if we can integrate the environment into SkyRL.🤯🤯🤯

A start toward real multimodality: an agent that can perceive, reason, and act in real time within open-world environments for hours. 🎬Project page: https://t.co/S8tyXNbRTv 📄Paper: https://t.co/Xu1Ysbw8Le More details: https://t.co/ZeaMFTwIrk Kudos to the team : )

Meet Slingshots // One. This inaugural batch includes leading-edge researchers advancing the science and practice of AI - with benchmarks, frameworks, and agents that ship real impact into the world. We're honored to support research from: @alexgshaw @Mike_A_Merrill

We are happy to release SkyRL tx 0.1 https://t.co/PSOuZciiGw, an open source unified training and inference engine that supports the Tinker API. This release has many performance enhancements and also new features but most importantly RL training is now working end-to-end. If you

SkyRL just crossed 1000 Github stars! Thank you to all the wonderful contributors and users building this project together 🥳 Check it out: https://t.co/CWlKue79JH

Honored and grateful to receive the Amazon AI Fellowship! Huge thanks to @AmazonScience for the support—excited for the journey ahead 💙💛