Blog
Post-Training Open-Weight Models on Legal Tasks with OpenRL
How we post-trained Gemma 4 and Qwen 3.5 on Harvey's Legal Agent Benchmark with OpenRL.
Delta Weight Sync in OpenRL
How OpenRL speeds up full-parameter RL post-training 5–6× by syncing only the weights that change as in-place vLLM patches.
OpenRL: Full Fine-Tuning, Multi-GPU, Agentic RL and More
A look at what we built in OpenRL over the last two months: full fine-tuning with delta weight sync, multi-GPU training with context parallelism, hero runs on dedicated GPUs, custom trainer images, the Harvey LAB recipe, and agent-friendly skills and observability.
Introducing OpenRL: A self-hosted post-training API for fine-tuning LLMs
Google Open Source Blog