Castform

@castformai prev @scifivc @stanfordsymsys

Castform is a post-training platform that lets engineering teams fine-tune and reinforcement-learn open-weights models on their own data, then export the trained weights and deploy them anywhere. The pitch behind today's open preview is that small, task-specific models can match or beat frontier APIs on the narrow jobs most teams actually run in production, at a fraction of the per-token cost. The target user is a working AI engineer who already has a RAG corpus, agent traces, or internal docs and would rather train a model they own than keep paying for general-purpose intelligence they do not need. The product handles the parts that usually make custom training painful. Castform auto-generates training data from existing corpuses and agent traces , with connectors for stores like Turbopuffer, Pinecone, Chroma, and Postgres on the retrieval side and Braintrust, Langfuse, and Langsmith on the traces side. Users define the environment and reward signals their agent is scored on, kick off a run, and get back a fine-tuned open model. The platform is self-service and pay-as-you-go, with free credits to trial a first training run , and when a stronger open-weights base model drops, the same recipe can be re-run to upgrade the task-specific model. The launch matters now because open-weights quality has caught up far enough that the cost gap with closed frontier models has become hard to ignore for teams burning meaningful inference spend. Castform is led by Girish (@googrish on X), who is building the company after stints at SciFi VC and Stanford's Symbolic Systems program, and is hiring across compute orchestration, training research, and developer storytelling.

Tags

Comments (9)