Jev

jeff is a self-hosted, MIT-licensed drop-in implementation of TypeSafe's jev System One API, powered by the GLiFormer encoder model. It accepts classification questions through a compatible SDK or the POST /v1/systemone endpoint and returns choice selections, ordered scores, and noul yes/no probabilities. The server supports CUDA, Apple MPS, CPU, and ONNX Runtime deployments, with batching, bearer-key authentication, rate limits, request limits, health and statistics endpoints, and deployment configurations for Modal. Its repository reports lower hosting costs than jev but weaker results on reasoning-heavy tasks.

kev is an open-source, local Jev-inspired decision model from Jared Palmer. It runs on a laptop and accepts a document or other state plus typed questions, returning calibrated probabilities rather than generated text. It supports yes/no (noul), multiple-choice, and ordered-score questions. The model uses a LoRA adapter and a readout head on a Qwen base model. It packs the state and all questions into one sequence, applies a block-causal mask so each question can see the state but not sibling questions, and performs one prefill pass without decoding. A pointer head scores each option against the question's decision token and applies softmax; the head is trained with cross-entropy on labelled outcomes. The local server exposes a TypeSafe System One-compatible API at POST /v1/systemone, allowing the official TypeSafe SDK to connect by changing its base URL. It also provides model information, option-order permutation, and separate-pass comparison endpoints, plus a local playground. The repository includes 0.5B, 0.6B, 4B, and 8B model checkpoints and recipes for local or Modal-based training and evaluation. Serving is tested on Apple Silicon, with larger checkpoints requiring bf16 on a 32 GB Mac. The server has no authentication and is intended for local use. The project is licensed under Apache-2.0; base models and datasets have separate licenses.

LocalJev is a TypeScript service that provides a local Jev-compatible decision API for applications using TypeSafe's typed decision interface. It exposes a POST /v1/systemone endpoint and translates typed Jev questions and application state into classification prompts for an OpenAI-compatible DiffusionGemma inference server. The service validates and retries malformed JSON, normalizes probability vectors, calculates Jev-compatible choices, expected scores, and entropy-based confidence, and returns the standard Jev response shape. Its probabilities are model-generated or self-reported rather than read directly from model logits, so the repository advises evaluating calibration for the intended workload. It runs with Bun and can use oMLX as the local inference server; the default LocalJev address is http://127.0.0.1:8080.

Jevlike is an open-source research starter for training small models that choose among a variable list of text options. Given a context and an option list, it represents each option as a query, attends to the context tokens to produce an option-specific context vector, applies a shared dot product, and runs a softmax across the options to return probabilities in one pass rather than generating an answer token by token. It provides a trainable byte-level encoder and an optional frozen Hugging Face encoder, with command-line tools for generating synthetic data, training, evaluation, and prediction from JSONL records. The repository also includes a visual option scorer used by Doom and chess controller examples. Training supports CPU, Apple MPS, and CUDA; the default encoder truncates contexts to 192 bytes and options to 32 bytes, and the project requires the complete option list at prediction time. Jevlike is an independent implementation rather than a copy of TypeSafe's commercial Jev model, and its code is released under the MIT License.

Simple Jev is an open-model classifier and scoring server from Featherless AI. It accepts shared context or chat messages plus questions, then reads the model's next-token logits for predefined answer labels instead of generating a prose or JSON completion. The server constructs JSON responses for choice classification, ordered rubric scores, and truth or support judgments, including confidence and label distributions. Its Hugging Face Transformers implementation identifies the exact common token prefix across questions, evaluates that prefix once, reuses its KV cache for question-specific suffixes, and batches the suffixes before scoring the allowed answer tokens. The response-scoring code normalizes the selected logits; shared context is counted once for usage purposes, and no output tokens are generated. The API provides non-streaming /v1/classifier and /v1/systemone endpoints, with local execution supported through Python, PyTorch, and Transformers. A public demo API is available, while production deployments can use Featherless paid plans or run the server themselves. The repository also includes RFDT, which fine-tunes models directly on allowed answer-token logits using the same prompt structure.