GitHub radar
LocalJev: Free local run of Jev typed classification
GitHub Next runs a local server on Jev: ask a question with multiple types, get responses back for all types in a single request, right on your Mac via DiffusionGemma without having to pay an API fee.
TypeSafe Jev’s API takes a message and a list of typed questions regarding that message (e.g. “what team should I assign this ticket to?”, “how urgent is this problem?”, “how frustrated is the customer?”) and returns answers for all the question types in one go. LocalJev implements the TypeSafe Jev protocol by mapping questions to a classification prompt, querying a local instance of DiffusionGemma through an oMLX server, and returning the answer in the same format as the paid TypeSafe Jev API. The main distinction between LocalJev and Jev is that while Jev returns probabilities in the model’s logit space, LocalJev prompts the model to return the probabilities in natural language, so the results are only approximations. Please note that it’s important to check the model’s calibration on your data before making critical decisions.
Why a vibe-coder should care
With LocalJev, you can classify messages on your Mac without paying for API calls — great for applications/agents that require messages to be classified (e.g., support ticket assignment, urgency estimation, sentiment analysis). LocalJev is suitable for straightforward and monotonous classification tasks where accuracy isn’t paramount. For high-stakes decision making, we suggest checking the model’s calibration on your data.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Deploy localjev from https://github.com/githubnext/localjev following the README — set up Bun, run an oMLX server with DiffusionGemma 26B-4bit, start localjev, and show me a test request classifying a support message with three questions: department, urgency, mood. Ask me if you need API keys.
Requires a Mac with 32 GB RAM for DiffusionGemma 26B 4-bit. Does not run on a 16 GB machine.
Open on GitHub▌ More finds