Baseten adds deferred tool loading for open models

Baseten announced support for deferred tool loading across DeepSeek, Kimi and GLM models on its Model APIs, allowing agents to retrieve tool definitions only when needed.
The feature supports Chat Completions and Messages API formats, according to Baseten's announcement. Instead of sending every registered tool's full description with each request, an agent begins with tool names and a search function. Relevant definitions are added to the conversation when requested, preserving the existing prompt cache.
Baseten tested GLM 5.3 on 55 tasks drawn from eight benchmarks. In its coding tests using Claude Code with three MCP servers, deferred loading reduced average cost by 62.5% and completion time by 26.2%. Those are company-reported results, not universal savings guarantees.
The company also reported that some individual benchmarks ran slightly slower, despite overall improvements. The practical benefit is reducing the cost of carrying large tool catalogs through every model request, without replacing the underlying model or agent harness.
Sources and citations
Revision history
Published .