Add bit-jev to Local / On-device Inference

[Zeaulo]

Project

• Name: bit-jev
• URL: https://github.com/Zeaulo/bit-jev
• Category: 3. Inference Engines & Serving → Local / On-device Inference (appended at the end of the subsection)

Why it belongs

bit-jev runs structured decisions (multi-choice, yes/no, ordinal score) locally on CPU. A pointer head on a distilled BitNet b1.58 backbone scores the options the caller declares, so no answer tokens are generated. The backbone ships as an I2_S GGUF (about 1.19 GB) with a float32 head. A native runner built on pinned BitNet/llama.cpp sources runs it, and a pip package (pip install bit-jev, python -m bit_jev.demo) downloads the model on first use.

Quality signals

• License: Apache-2.0 for code. The model weights have no separate open-weights license; the model card documents data provenance (Yelp permission pending as of 2026-09-28).
• Maintenance status: active (last push 2026-09-30; PyPI 0.13.13).
• Documentation/examples: README (zh/en), CPU quick start, GGUF package guide, benchmark protocol with per-run JSON records.
• Distinction from similar projects: unlike general local LLM runtimes, it is a task-specific decision model that returns option scores/probabilities. Microsoft BitNet (already listed) is the backbone and inference base.

python3 tools/validate_awesome.py --skip-remote: 0 errors; 10 pre-existing duplicate-repo warnings, none from this entry.

Disclosure: I maintain bit-jev; this PR was prepared with AI assistance.