Baseten made inference for pyannote's speaker diarization model 9.6x faster. The blog also covers delta weight synchronization optimizations for hosted RL rollouts, Baseten leading in the Coval voice AI benchmark, and new MCP and skills for coding agents to use Baseten.
Baseten announced a partnership with OpenAI, and together with the NVIDIA OpenShell and Blaxel sandboxes, safeguards the open frontier. Baseten launched Hosted Tools and supports Baseten Loops fine-tuning based on LangSmith traces; LangChain also uses Baseten Loops to train custom models for the LangSmith Engine. In addition, Baseten released DeepSeek-V4.1-Flash, providing more efficient prefill for coding agents.