kev vs strands-decider
Open reproduction of TypeSafe's Jev: a LoRA + readout head on Qwen that answers many typed questions about one document in a single prefill pass, returning calibrated probabilities. — versus — Strands' small decision model (2B, Qwen3.5-based): pick between options, answer yes/no or rate on a scale with calibrated confidence. CLI and /v1/systemone server on CUDA, MPS, MLX or CPU.
Both answer many typed questions about one state in a single pass with calibrated probabilities; kev reproduces Jev as a LoRA head on Qwen, strands-decider ships its own trained 2B checkpoint.
| kev | strands-decider | |
|---|---|---|
| Stars | 8.7k | 579 |
| Forks | 571 | 77 |
| Language | Python | Python |
| License | Apache-2.0 | Apache-2.0 |
| Last activity | 5 days ago | 2 days ago |
| Topics | decision-models, training, local | decision-models, local |
| Curated connections | 12 | 5 |
kev — the curator's take
Reach for kev when the task is decisions, not prose: route this ticket, score this risk, answer 30 yes/no questions about one document — and you want a number you can threshold on rather than text you have to parse. The head is trained with cross-entropy on labelled outcomes, so the probabilities mean something (kev-8b: Brier 0.34, 8% confident errors out of domain), and a block-causal mask keeps questions from seeing each other, verified to 4e-6 against separate requests. It speaks TypeSafe's System One API, so their SDK works against localhost with a base_url change. Not a general chat or agent model - it never decodes, and it only answers the question types you define (noul/choice/score). Don't use it zero-shot on your own domain either: the value is in training the head on your labels, and the three 0.6B/4B/8B checkpoints are explicitly previews that failed the author's own release screen on held-out rule reasoning. Real Jev still wins out of domain (0.86 vs 0.77) - kev's pitch is that it runs on your laptop and you own the weights.
strands-decider — the curator's take
Use it for the many small judgments inside an agent workflow — routing, triage, urgency, sentiment — where an LLM call is too slow and training a classifier too much work; answers at ≥0.9 confidence were right ~95% of the time on unseen tasks, so threshold and escalate the rest. It decides, never writes. At 2B it runs on a laptop; for agent-action ranking at scale, CLM's cached-embedding design is faster.