ByteDance & Tsinghua Release CUDA-Agent: RL-Based High-Performance Kernel Generation
AIFlow_ML · x · 2026-08-17
ByteDance and Tsinghua University SIA have released CUDA-Agent, a system utilizing large-scale Agentic Reinforcement Learning (RL) for high-performance CUDA kernel generation.
Key Highlights:
- Performance: Achieves SOTA on KernelBench, consistently outperforming Claude Opus-4.6 and Gemini 3 Pro, with significant gains on the most difficult cases.
- Dataset: Released CUDA-Agent-Ops-6K (6,000 samples), built by fusing operators from PyTorch/Transformers using an LLM and filtering for executable, deterministic, and non-trivial samples.
- Open Source: The project includes the training data, expert-designed SKILL.md, and the agent environment to support community research in LLM-based CUDA generation.
More from Infra
- CoreWeave's debt costs falling: DDTL spreads tightened from S+9.62% to S+2.25% — BenBajarin · 2026-10-16
- SoftBank lines up US IPOs for SB Energy and AI robotics spinoff Roze, valuations top $50B — nordicinst · 2026-10-16
- TRL reproduces Modal's mixed-precision async RL: plain GRPO collapses, SC+MIS holds 0.72 — QGallouedec · 2026-10-16
- Hugging Face scales Indic pretraining with 2.8T tokens and an open data pipeline — Hugging Face Blog · 2026-10-16
- NVIDIA Vera Rubin NVL72 hits up to 5.7x GB300 throughput per MW on agent benchmark — ArtificialAnlys · 2026-10-16
- 14 open-source LLM observability tools to know in 2026, from Langfuse to Laminar — TheTuringPost · 2026-10-16