Open-source Kev-0.5B: a tiny Qwen2.5-0.5B decision model you can train on a MacBook
sull · x · 2026-09-19
jaredpalmer released Kev-0.5B, a tiny Jev-like decision model built on Qwen2.5-0.5B, with weights on GitHub (121 stars already).
- LoRA adapter plus a readout head on Qwen2.5-0.5B, exposed via a TypeSafe-compatible API
- Packs a document and many typed questions into one sequence; a block-causal mask lets each question attend to the document but not other questions, so all calibrated probabilities come out in a single prefill pass with no decoding
- A pointer head scores each question's options and softmaxes them; the head is trained with cross-entropy against labeled outcomes, so probabilities are learned rather than generated
- Trainable and runnable on a MacBook Pro
More from coding & agent
- Human Oversight Is 70-75% of AI Agent Costs in Banking, McKinsey Finds — mikeflache · 2026-11-04
- Over 80% of Major Financial Institutions Now Run Agentic AI in Production — mikeflache · 2026-11-04
- Mnemos MCP nears launch after shared memory layer fix — RileyRalmuto · 2026-11-04
- Moving mobile dev to remote machines makes it multiplayer across product, design, engineering — rudrank · 2026-11-04
- Kuaishou's EvoRec: self-evolving agentic recommender lifts ad revenue 4.6% — _reachsumit · 2026-11-04
- TRIO: three specialized agents with aligned objectives refine retrieval-augmented generation — _reachsumit · 2026-11-04