知识库收录材料 · 原始文件 ↗ · 来源、版本与许可

v1.2.8 additions: decider-35b-a3b and reflex-27b

Both systems ran the unchanged frozen JevBench v1.2 set: 534 decisions, including held-out items, one request at a time through the existing TypeSafe adapter. No task, answer, scoring rule or earlier row changed.

decider-35b-a3b: public ungated Apache-2.0 checkpoint Mapika/decider-35b-a3b at revision 30c5f9880f9f9eef56ef55e1b44aa6e61d3b7919; Apache-2.0 server code at commit c4daaac28af9fea95d627015cffa2dd5a5926ee6. The server returns its native option probabilities.
reflex-27b: MIT server code at commit b77f03bab8e00d7319e64550e062c1eac8111875; public ungated Apache-2.0 Qwen/Qwen3.8-27B checkpoint at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. The requested configuration uses no adapter or fitted calibration file, temperature 1, and averages two option orders.

Both ran serially on a disposable RunPod H100 NVL (96 GB) reached from Sandy over the internet. The pod contained no API keys or tokens. Self-host latency uses the standing ×2 + 0.15 s adjustment. Estimated cost uses the closest public hosted input tariff: $0.10/M input for a 35B-A3B model and $0.214/M input for the exact Qwen3.8-27B base; neither gets an automatic free-tier Cost score.