Kev-0.5B
#10 in Lütte & Edge-ModelleUnknown · 0.5b · 3× · tolest 20. Sept. 2026
48
Momentum
Kev-0.5B is an Open Source, very small decision model by developer Jared Palmer, built on the base model Qwen2.5-0.5B. Instead of generating text, it processes a document together with multiple typed questions in a single forward pass and delivers calibrated probability distributions as answers. It is designed as a LoRA adapter plus small pointer head, offers a TypeSafe-compatible API, and can be fully trained and run locally on a MacBook Pro (Apple Silicon). Weights and model card are freely available on GitHub and Hugging Face.
Momentum-Verloop
22.06.20.09.
Features
| Key Benchmark (%) | 79.9% accuracy on held-out split of the six training sources (1,350 questions, self-reported) |
| License | Apache-2.0 (base, adapter and head) |
| Multimodality | Not multimodal; text-only processing, does not generate text but returns probability distributions for typed questions |
| Platform | Runs locally on MacBook Pro (Apple Silicon, tested on MPS); training also via cloud (H100 via Modal) |
| Release Date | Not officially documented with exact date; announced 2026 via X/Twitter and GitHub release v0.1.0 |