Organization repository on GitHub
TypeLLM: LLMs with type-safe generation
- Forks
- 61
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
TypeLLM
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
- Stars
- 941
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- TypeLLM
- Last verified
- Oct 2026
About this project
Homepage • Blog • Docs • Examples • API • Contact 🚀 TypeLLM API is live: typed outputs without serving a model, $5 free credit. Try it · Quick start · Examples. [2026/10/08] Conditional fields can test how sure an answer is: "when": {"category": {"confidence": {"gte": 0.8}}}. [2026/10/05] Added objects and arrays: group typed properties into one value, or return a variable number of typed items.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Declares string, integer, number, boolean, and enum fields with instructions, including nullable types and bounded numeric decoding.
- Runs independent fields together and honors depends_on graphs with shared-prefix reuse, plus conditional fields answered only when triggered.
- Offers per-field thinking with a token budget alongside image inputs for vision-language models and permutation averaging for order bias.
- Uses one client for cloud or self-hosted models, with typed results, per-field reasoning traces, and token usage on every response.
Watch out
Free text stops after a bounded token limit unless raised explicitly, enums cap at 24 values, and thinking mode adds tokens. Cloud use needs an API key; self-hosting needs a GPU server. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Jev answers only decisions in one non-autoregressive pass with confidence; TypeLLM keeps native thinking and free-form generation while constraining outputs to JSON Schema, with dependency graphs and image inputs.
Categorical picks cost about one token and numeric decoding is bounded, with shared-prefix KV reuse; only the fields you flag for thinking pay reasoning tokens, and they reason side by side.
Ask for length explicitly and raise the text token cap, mark fields nullable with null in the type and enum, and adjust the numeric digit cap; omitted instructions fall back to description or field name.
No. Start with the hosted API key and playground, then point the same client at a local SGLang endpoint and model name when you want to serve a compatible checkpoint yourself.
String, integer, number, boolean, enum choice, object and array. Enum choices support string, integer and number with 2 to 26 values; free text stops after 128 tokens by default unless you raise text_max_tokens, and numeric answers use plain decimal notation with at most 32 digits unless you change numeric_max_digits.
pip install -U typellm, then either sign in and create a key for the hosted TypeLLM API (set TYPELLM_API_KEY, with $5 free credit) or point TypeLLMClient at an SGLang server you run, such as http://127.0.0.1:30000 with Qwen/Qwen3.8-27B. Both clients take the same requests.
On the 231 public JevBench tasks it scored 195/231 without thinking and 228/231 with thinking. The full per-task answers and the method and configuration are published, and the project says choices stay within the allowed values with negligible output-token cost beyond optional thinking.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPython164GitHub stars
End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zwliJay/jev-forgePython
zwliJay/jev-forgePython70GitHub stars
Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bladedevoff/stuntdPython
bladedevoff/stuntdPython42GitHub stars
Open harness that renders allowed JSON answers on a diffusion canvas, denoises for a fixed step budget, then picks the top allowed tokens from final logits with no training.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
JoshuaSP/open-jevPython
JoshuaSP/open-jevPython16GitHub stars
Python kit pairing a Jev CLI for verify, screen, classify, extract, match, route, ask, find, rerank, compact, and batch with an MCP server exposing the Jev ruleset over stdio.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
utk2103/jev-studioPython
utk2103/jev-studioPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.