GitHub radar
SemIf: Semantic AI Decisions without Text Generation
Open source standalone tool to query typed options from a local model without needing to generate an answer sentence or parse jsons (5.21x faster than chat models).
SemIf is an open source standalone tool that mimics the interface of TypeSafe’s Jev by allowing you to query typed options from a local model rather than generating an answer sentence. Given some text context and a set of questions (each question is a multiple choice or a boolean), it will output the probability of each option. There are no answer tokens generated or jsons parsed. With an RTX 3090, it takes 1.023s to make 21 binary decisions whereas it takes 5.332s to generate an autoregressive json array (5.21x faster). Supports Apple Silicon through the MLX backend with shared state parallelization and serial prefix reuse. Also has a webgpu browser demo that doesn’t require installation. Was previously known as OpenJev but has nothing to do with TypeSafe.
Why a vibe-coder should care
If your agent needs to classify messages, ticket them, or evaluate predicates, then SemIf lets you skip a chat model query and go straight to querying the probabilities which should be faster and cheaper. All local, no cloud API necessary. For basic classification tasks, it achieves similar performance to Jev on all the tasks I tested.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Look at SemIf: https://github.com/TheoLeeCJ/SemIf — a tool for text classification without generation. Install it on my Mac using the MLX backend from the README, run the example from examples/decisions.jsonl and show me the results; ask me if you need a model path.
Runs on any Apple Silicon Mac via the MLX backend — no discrete GPU needed. On Windows or Linux, an NVIDIA GPU with 8+ GB VRAM is required for the 4B model.
Open on GitHub▌ More finds