GPT-6 made computer use the frontier. We make it open. 🚀 Meet OpenGCUA, an ongoing project to open the full stack for generalist computer use. All 8 frontier models improve on OSWorld-v2, GPT-6 Astra included. On average: +19.6 points and 53% fewer steps than batched GUI-only. The idea: let the model decide how to observe, act and organize its work, across visual and programmatic interfaces. We call such a system a generalist computer-use agent (GCUA). Two designs make it work: ⚙️ Code execution: code becomes the model's only action space. Each step is one program that can combine GUI actions, shell commands and app APIs, with loops and checks 🧭 Harness-to-model decision transfer: As models become more capable, we envision progressively shifting selected workflow decisions from fixed harness rules to the model Open today: 📄 Tech report 🧠 OpenGCUA harness and prompts, with an OSWorld integration example 📜 Evaluation trajectories of all 8 models 💻 macOS real-device harness, as a native menu-bar app (preliminary) 🔌 Cross-Device SDK (preliminary) 🧪 7 interactive labs and real-Mac demos Paper: arxiv.org/abs/2610.22200 Code: github.com/Tongyi-MAI/OpenGC… Labs and demos: tongyi-mai.github.io/OpenGCU… A closer look below 🧵

Introducing 🔥CodeT5+🔥, a new family of open-source code LLMs for both code understanding and generation, achieved new SoTA code generation performance on HumanEval, surpassing all the open-source code LLMs. Paper: arxiv.org/pdf/2305.07922.pdf Code: github.com/salesforce/CodeT5… (1/n)

CodeRL advances program synthesis by integrating pretrained language models + deep reinforcement learning. Using unit test feedback in model training and inference + an improved CodeT5 model, it achieves SOTA results on competition-level programming tasks. blog.salesforceairesearch.co…