JevK5 open decision model ranks #2 of 76 on JevBench, outputs calibrated probabilities in ~13ms
airesearch12 · x · 2026-09-24
Developer alibiserikbay open-sourced JevK5 under Apache-2.0: a decision model built on Qwen3.5-4B with a distilled LoRA. Given a state (ticket, log, policy, diff) plus typed questions, it returns calibrated probabilities for every option in a single forward pass — zero generated tokens, 13 ms on one GPU.
Key points:
- Ranked #2 of 76 systems on JevBench v1.4 (62.04), behind only Jev 1.13.0 (63.29), ahead of every other open system
- Judge tier 0.945, identical to Jev's; fastest Speed axis in the top five
- Scores measured by the benchmark's author on their own hardware
- Ships 4B weights, CUDA graph runtime, and training code; compatible with TypeSafe's /v1/systemone API
More from coding & agent
- Every hands its September Brooklyn meetup entirely to an AI agent, from menu to guest list — danshipper · 2026-09-24
- Designing daily agents: decide upfront what happens after a missed run — gethackteam · 2026-09-24
- Running a brand with only AI agents: the Notch experiment applied to nail polish ads — azed_ai · 2026-09-24
- "The replacement you trained just became your boss" — Codex gets StackOverflow plugin — cto_junior · 2026-09-24
- Splitting sandbox base layers with Nix: fewer images, more auditable agent environments — sloppenheimer · 2026-09-24
- Devin adds native Teams support and first-party Microsoft 365 integration — DevinAI · 2026-09-24