TypeLLM: Open-Source "Workflow Model" Runs Your Entire Pipeline in One API Call

rohanpaul_ai · x · 2026-10-15

Open-source project TypeLLM pitches a "workflow model": chat models return a message and decision models return a choice, but TypeLLM takes an entire workflow — multiple typed answers, fields depending on others, conditional fields — as a single JSON Schema and executes it as a dependency graph over one shared context in a single API call.

Under the hood, TypeLLM sits on SGLang and compiles JSON Schema into constrained decoding: one scored token for enums/booleans, bounded digits for numbers, with the dependency graph reusing cached prefixes. Fields structurally cannot return out-of-schema values — schema-guaranteed outputs from a stock Qwen checkpoint without touching its weights.

Related event: TypeLLM Open-Sources Workflow Models: Whole Workflow in One API Call(2 posts)→

Original post →

More from coding & agent

coding & agent channel →