LLM2Jev
Independent Apache-2.0 server: local text and vision models expose POST /v1/systemone from prefill logits on SGLang, Transformers, or MLX.
Overview
LLM2Jev (github.com/Yinsongxu/LLM2Jev, Apache-2.0) from Yinsong Xu is an independent open-source project. It is not affiliated with or endorsed by Jev or TypeSafe. The stack turns compatible causal and vision language models into Jev-style structured decision outputs by reading probabilities from logits during prefill, without token-by-token decoding. Backends include SGLang on NVIDIA Linux, Hugging Face Transformers, and MLX on Apple Silicon. The repo ships a compatible POST /v1/systemone HTTP service, a Python API, multimodal text and image inputs, and staged candidate scoring for SGLang Radix Cache prefix reuse. Web and Snake demos and a Valen Sokoban example added October 8 are included. docs/jevbench.md documents 231 public JevBench items with reproduction commands.
Problem: Teams want System One shaped HTTP on GPUs they already own, but hosted Jev bills per call and chat-style JSON bridges add decode latency.
Built for: Engineers running local text or vision language models who need POST /v1/systemone compatibility, multimodal state, and logits built during prefill only.
First indexed on Jev Directory: 2026-09-19
Creator and team
- Name
- Yinsong Xu
- Handle
- @Yinsongxu
How Jev is used
- Role in the product flow
- Local inference server and Python scorer: shared state plus Choice, Score, and Noul questions become one or staged prefill passes over hub weights
- Primitives
- ChoiceScoreNoul
- State in
- System One compatible JSON: shared state strings, optional images in state or instructions, and per-question criteria lists per docs/usage.md and docs/multimodal.md.
- Decision out
- Probability maps assembled from logits during prefill; HTTP responses match /v1/systemone envelope expectations for SDK and JevBench adapters.
- Compile each question into model prompts per docs/request-to-model.md
- Score criteria tokens from logits in one pass or staged Radix-friendly batches on SGLang
- Optional python -m jevbench.cli run against a local endpoint for public task files
- Serve with the documented online HTTP service on a chosen port
LLM2Jev is a self-hosted wire-compatible server: you keep weights on SGLang, Transformers, or MLX and expose the same POST shape agents already use. featherless-simple-jev also assembles JSON from logits but targets Featherless or a HF classifier surface, not a full /v1/systemone port with multimodal state and JevBench CLI. nokia-applied-research-anyjev wraps vLLM with calibration levels and BANKING77 tables; LLM2Jev focuses on prefill-only speed, Radix staged candidates, and Mac MLX paths. liuziyu77-valen trains multimodal decision checkpoints; LLM2Jev's Valen Sokoban example scores single-step Choice on project eval data without shipping Valen weights. LLM2Jev is independent of TypeSafe. Test the numbers on your own hardware before you switch production traffic.
Sourced performance claims
- README JevBench section cites 76.2% accuracy on 231 public items with LLM2Jev and Qwen3.5-4B, P50 latency 48 ms and P95 346 ms on the local endpoint versus Jev 1.13.0 P50 652 ms and P95 722 ms in the published comparison table.Source: github.com/Yinsongxu/LLM2Jev README and docs/jevbench.md
- docs/jevbench.md lists 231 public items across original, easy, and hard datasets used for the accuracy and latency summary.Source: github.com/Yinsongxu/LLM2Jev docs/jevbench.md
- Public GitHub repo Yinsongxu/LLM2Jev had 417 stars and 45 forks as of October 9, 2026.Source: GitHub API, 2026-10-09
Features and stack
Features
- POST /v1/systemone compatible HTTP service documented in docs/usage.md
- SGLang, Transformers, and MLX-VLM backends with multimodal inputs
- Prefill-only logits scoring without autoregressive answer decoding
- Staged candidate submission for SGLang Radix Cache on cold requests
- Web demo, Snake demo, MuJoCo pick-and-place, and Valen Sokoban examples
- jevbench CLI for 231 public items with typesafe adapter to a local endpoint
Stack
- Python 3.12+
- SGLang
- Hugging Face Transformers
- MLX on Apple Silicon
- uv or pip install extras
Pricing: Apache-2.0 open source; you pay for GPUs, electricity, and model weights you download.
Links
FAQ
- Is LLM2Jev affiliated with TypeSafe or Jev?
- No. The README banner states LLM2Jev is an independent open-source project not affiliated with or endorsed by Jev or TypeSafe.
- How is LLM2Jev different from Simple Jev?
- Simple Jev exposes /v1/classifier from Featherless or your HF server. LLM2Jev targets full /v1/systemone compatibility, multimodal state, MLX on Mac, and ships JevBench reproduction docs.
- How is LLM2Jev different from AnyJev?
- AnyJev adds L0 through L2 calibration heads on vLLM hidden states. LLM2Jev scores logits during prefill on SGLang, Transformers, or MLX without Nokia's pipeline levels.
- Can I run it on a Mac without NVIDIA?
- Yes. docs/usage.md documents an MLX backend on Apple Silicon for text and image scoring plus a compatible HTTP service.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Where to run Jev
Compare official TypeSafe, Vercel AI Gateway, OpenRouter, Cloudflare Workers AI, and classifier.dev with a fact table.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.