Mubit launches a platform that trains small task models from agent calls, and the open model Banyan Sapling

Mubit has launched a platform for teams to train their own models for repetitive agent tasks, along with an open research-preview model and a new benchmark, according to TestingCatalog's account of the founder's launch thread.

• Platform: Mubit sits as a proxy base URL in front of OpenAI, Anthropic or Gemini API calls and trains a small model on the agent's tasks. Mubit says the small model takes over a task once it matches the frontier model in Mubit's own tests. It can also learn from past calls imported from Langfuse, Braintrust or Pydantic Logfire.
• Banyan Sapling: a research-preview open model under the Apache 2.0 license. It judges whether a given context supports a claim, violates it, lacks the information, or contradicts it, and gives calibrated probabilities. In Mubit's evaluation it answered 779 of 1,731 cases correctly and marked the rest "unsure".

The performance figures are Mubit's own.