TypeLLM releases open-source JSON Schema workflow engine built on SGLang

TypeLLM, an open-source project, releases an engine that takes a JSON Schema describing a multi-step workflow and executes it in one API call.

The project says it runs the schema as a dependency graph over a single shared context, using constrained decoding on SGLang so enum and boolean fields return only allowed values, numbers stay within bounded digits, and cached prefixes are reused.

Rohan Paul, posting on X, reports JevBench results of 84.42% accuracy without thinking and 98.70% with thinking for TypeLLM running on Qwen3.8-27B. He says the checkpoint was used without fine-tuning. These figures come from the post and have not been independently verified.