LocalJev
GitHub Next Bun server: local POST /v1/systemone over oMLX DiffusionGemma chat JSON, TypeSafe SDK compatible, README honest on calibration.
Overview
LocalJev (github.com/githubnext/localjev, MIT) is GitHub Next's TypeScript bridge that exposes POST /v1/systemone on port 8080 by default. It translates shared state plus Choice, Score, and Noul questions into classification prompts, asks an OpenAI-compatible chat endpoint (defaults: upstream http://127.0.0.1:8000, model diffusiongemma-26B-A4B-it-4bit), validates JSON probability output with retries, normalizes vectors, and returns Jev-compatible envelopes. README is explicit: wire-compatible with System One, not logit-equivalent to OpenJev's one-step structured read. GET /ready checks upstream model availability; jev-latest and jev-preview aliases satisfy SDK defaults.
Problem: Mac builders want System One-shaped HTTP without shipping screenshots to TypeSafe, but DiffusionGemma on oMLX lacks OpenJev's structured logit read primitives.
Built for: Developers running Bun 1.2+, oMLX with diffusiongemma-26B-A4B-it-4bit, and TypeSafe SDK clients pointed at localhost.
First indexed on Jev Directory: 2026-09-25
Creator and team
- Name
- GitHub Next
- Handle
- @githubnext
How Jev is used
- Role in the product flow
- Local Bun server maps System One JSON to chat prompts and back with schema validation
- Primitives
- ChoiceScoreNoul
- State in
- Standard System One body: model alias, state string, and questions map with choice, score, or noul criteria per README curl.
- Decision out
- Jev-shaped choices, expected scores, and entropy-based confidence from model-reported JSON probabilities after normalize and retry logic.
- Run oMLX (or other OpenAI-compatible server) with configured DiffusionGemma checkpoint
- bun install, copy .env, set LOCALJEV_UPSTREAM_API_KEY, bun run start on :8080
- curl /ready then POST /v1/systemone or point typesafe_sdk at TYPESAFE_BASE_URL
- Tune LOCALJEV_QUESTIONS_PER_CALL and MALFORMED_RETRIES for your workload calibration
LocalJev is the honest Mac compromise: keep TypeSafe SDK code, accept that probabilities came from a prompted JSON scalar instead of OpenJev's diffusion read. featherless-simple-jev and jaredpalmer-kev chase logit-native servers; githubnext-localjev meets oMLX where it is today. theoleecj-semif and wfzyx-von are alternative open stacks if you can leave DiffusionGemma. nokia-applied-research-anyjev is for vLLM operators measuring ECE; LocalJev is for curl-friendly localhost demos with documented limitations. Read docs/evaluation-results-2026-09-18.md before you trust bake-off rankings on your ticket queue.
Sourced performance claims
- README states probabilities are generated or self-reported by the model, not read from logits, and recommends evaluating calibration before consequential gates.Source: github.com/githubnext/localjev README Why a bridge is needed
- Defaults document upstream :8000, LocalJev :8080, LOCALJEV_MAX_INFLIGHT 2, LOCALJEV_MALFORMED_RETRIES 2, and chunking limits 16 questions / 128 outcomes per call.Source: github.com/githubnext/localjev README Configuration table
- Public GitHub repo githubnext/localjev had 771 stars and 49 forks when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- POST /v1/systemone and /ready on Bun with .env driven config
- TypeSafe Python SDK works with TYPESAFE_BASE_URL and dummy API key
- Malformed JSON retries and per-call question chunking
- OpenJev vs oMLX gap explained with LM Studio status notes in README
- Evaluation guide and September 2026 bake-off report in docs/
Stack
- TypeScript
- Bun 1.2+
- oMLX OpenAI-compatible chat
- DiffusionGemma
- typesafe_sdk client
Pricing: Open source MIT server; inference cost is your local oMLX GPU time, not TypeSafe tokens.
Links
FAQ
- Is LocalJev the same as OpenJev?
- No. README contrasts OpenJev's patched vLLM structured read with this chat-prompt bridge. Expect similar JSON shape, different probability semantics.
- Which upstream server works?
- Defaults target oMLX on port 8000. Any OpenAI-compatible chat API works if you set LOCALJEV_UPSTREAM and model env vars.
- Can I use the TypeSafe SDK?
- Yes. README shows TypeSafeClient with TYPESAFE_BASE_URL http://127.0.0.1:8080 and any API key unless LOCALJEV_API_KEY is set.
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.
- 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.