TypeLLM opens a hosted playground with $5 credit; API goes to early access
TypeLLM, the SGLang-based runtime that forces existing LLMs into schema-typed answers, now runs as a hosted service. “Not another Jev,” the team posted on 29 September at 11:00 BRT: the playground is “open to all with $5 credit”, and the API is “rolling out to early-access users in the coming days”.
What you get. Typed outputs in more shapes than a decision model offers (string, integer, number, boolean, enum), plus image input, dependent fields (depends_on) and an optional thinking mode. The hosted API is POST https://api.typellm.ai/v1/generate, with a Python client (pip install -U typellm). API keys need an early-access code for now.
Pricing (from typellm.ai). Input tokens $0.05 per million (context, questions and images). Thinking tokens $0.50 per million, charged only for questions that set "thinking": true. Typed outputs are free. New users get $5.
Benchmark (UnverifiedClaim). On 231 public JevBench tasks, TypeLLM’s site reports 84.42% for TypeLLM on Qwen3.8-27B without thinking and 98.70% with it. The same table lists Jev 1.13.0 at 86.58%. These are TypeLLM’s own runs and match the figures already on our page (195/231 and 228/231).
How we list it. Nothing changes in the classification: TypeLLM is still constrained decoding over existing models, with no decision weights of its own, so it stays under runtimes and off the quadrant. The runtime page now has a “Hosted service” section.