jevos
feder-cr/jev CPU binary: OpenVINO int8 jevos-v4 serves POST /v1/systemone on localhost in tens of milliseconds.
Overview
jevos (github.com/feder-cr/jev, feder-cr/jev on Hugging Face) from feder-cr and Loris Salsi (@LosaLosSantos) is a single C++ binary that serves POST /v1/systemone in TypeSafe Jev wire format on CPU via OpenVINO int8 weights (jevos-v4 release). README quickstart: extract platform archive, unzip jevos-v4-openvino-int8.zip into model/, run ./jev serve on port 8017 by default. Supports parallel noul, choice, and score questions on one state JSON; GET /health reports readiness and model SHA-256; ./jev decide request.json answers offline. Latency table on README cites median 28 ms short and 130 ms long requests on Intel Core Ultra 7 255H versus hosted Jev and Laya on the same six-task accuracy grid. Optional JEV_API_KEY bearer auth; about 1 GB RAM. GGUF builds for llama.cpp ship beside OpenVINO in releases.
Problem: Teams want Jev-shaped noul, choice, and score on a laptop CPU without cloud tokens, Python runtimes, or a separate GPU server.
Built for: Self-hosters and EU privacy-minded engineers who can download release binaries or Hugging Face weights and point existing System One HTTP clients at localhost.
First indexed on Jev Directory: 2026-10-15
Creator and team
- Name
- feder-cr
- Handle
- @feder-cr
“Decisions on a laptop CPU in 28 to 130 ms. Send a text and a question and get back the probability of each answer, in TypeSafe Jev's API, on your own machine.”
How Jev is used
- Role in the product flow
- Local CPU System One server: parallel noul, choice, and score on one state read per request
- Primitives
- ChoiceScoreNoul
- State in
- JSON state string or object plus questions map with type noul, choice, or score per README curl and API section.
- Decision out
- Wire-compatible answers with probabilities; choice and score include per-option probabilities and confidence for human escalation; usage reports input_tokens with zero output_tokens.
- Download jev-linux-x64 or Windows zip plus jevos-v4-openvino-int8.zip from GitHub releases
- Place model/ beside binary or hf download feder-cr/jev model/* into jev/model/
- Run ./jev serve with optional --threads, --host, --port, --model-dir
- POST /v1/systemone from curl or TypeSafe-compatible SDKs
- Optional bearer JEV_API_KEY on all routes except /health
jevos is a self-contained CPU server for teams that want TypeSafe wire format without a Python or Node runtime at serve time. receptron-laya and jaredpalmer-kev need those stacks and larger downloads; wfzyx-von targets sub-25 ms encoder inference with a different training story; yinsongxu-llm2jev wraps your own Qwen weights behind /v1/systemone. jevos ships pinned int8 OpenVINO weights plus a GitHub wiki co-maintained with Loris Salsi. theoleecj-semif documents another open server path when you want research flexibility over a single release binary. README states code written for Jev SDK yes or no paths works unchanged; check score maturity notes (early) before production rubrics.
Sourced performance claims
- README benchmark chart: jevos-v4 median latency 28 ms short and 130 ms long on Intel Core Ultra 7 255H 16 threads versus Jev about 311 to 314 ms and Laya about 129 to 480 ms on the same harness.Source: github.com/feder-cr/jev README Benchmarks
- README feature table: jevos-v4 supports yes or no, multiple choice, and scores (scores marked early), runs on your machine, free cost, 8,192 token context.Source: github.com/feder-cr/jev README Benchmarks table
- Public GitHub repo feder-cr/jev had 1,325 stars and 159 forks as of October 15, 2026.Source: GitHub API, 2026-10-15
Features and stack
Features
- Single binary CPU inference with OpenVINO int8 jevos-v4 weights
- POST /v1/systemone compatible with TypeSafe Jev wire format
- Release archives for Linux and Windows plus HF model download
- Latency and six-task accuracy table versus Jev, Laya, and Qwen3.5-4B on README
- Optional JEV_API_KEY, ./jev decide offline mode, GGUF for llama.cpp
- GitHub wiki for deeper guides and measurements
Stack
- C++
- OpenVINO int8
- CPU only (no Python runtime required to serve)
- Hugging Face feder-cr/jev weights
Pricing: Free open weights and binaries; you pay for the machine and electricity. No per-token cloud bill when self-hosted.
Links
FAQ
- Is the GitHub repo name jev or jevos?
- Repository is feder-cr/jev; product and release tag use jevos (for example jevos-v4-openvino-int8.zip). This directory slug is feder-cr-jevos with title jevos.
- Does jevos need Python?
- README quickstart serves with ./jev only. Python appears in Build from source for OpenVINO SDK, CMake, and tests, not for the default serve path.
- How does jevos compare to Kev or Laya?
- README table contrasts CPU binary jevos-v4 against cloud Jev and local Laya on latency and six benchmark tasks. Kev and Laya listings document Python SDK and HF training paths; jevos optimizes one int8 CPU binary.
- Can I use the TypeSafe SDK?
- README API section states POST /v1/systemone matches TypeSafe Jev wire format so SDK code for yes or no questions can target localhost:8017.
- Where do weights live?
- GitHub release jevos-v4-openvino-int8.zip or hf download feder-cr/jev --include model/* per README quickstart.
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.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.