Jeeves：带推理的开源 Jev 风格决策模型（PostHog） Jeeves: open Jev-style decision model with reasoning (PostHog)

来源Sources

为什么推荐Why we recommend it 收录理由：晨间已收 Jeff（本机小决策模型），晚间 HN 出现 PostHog 开源、带推理的 Jev 风格 9B，补上 System One 线的「会想」一档。边界：基准与延迟是 README 自述 / Kev 公布对照，本库未复跑；需 CUDA GPU，未代装。Inspired by Kev，非 TypeSafe 官方。 Why: morning already catalogued Jeff (tiny local decision models); evening HN brought PostHog’s open reasoning Jev-style 9B — the “thinks first” tier on the System One line. Boundary: benchmarks and latency are README / Kev-published claims, not re-run here; CUDA GPU required, not installed for you. Inspired by Kev; not TypeSafe official.

要点（官方称 / 社区称）Highlights (vendor / community claims)

• README：9B Jev-like 模型（Qwen3.5-9B LoRA + pointer head），先推理再决策；带 diffusion drafter；训练代码与 train/dev/test 数据一并开源。 README: a 9B Jev-like model (Qwen3.5-9B LoRA + pointer head) that thinks before it decides; ships a diffusion drafter; training code and train/dev/test data are open.
• 作者基准表（自述，非本库）：Test overall 0.889（Kev-9B 0.822 / Jev 0.857）；JevBench overall 0.935（Jev 0.866）；hard 0.865（Jev 0.730）。无 thinking 同 checkpoint 在作者 test split 上 0.804（有 thinking 0.840）。 Author benchmark table (claim, not this library): Test overall 0.889 (Kev-9B 0.822 / Jev 0.857); JevBench overall 0.935 (Jev 0.866); hard 0.865 (Jev 0.730). Without thinking the same checkpoint scores 0.804 on the author’s test split (0.840 with thinking).
• API 与 Jev 同形：`python -m inference.serve` 后打 `/v1/systemone`；支持 choice / noul / score。延迟自述：无 thinking ~0.3 s，有 thinking 中位 ~3.3 s（H100）；需 CUDA。 Jev-shaped API: after `python -m inference.serve`, hit `/v1/systemone`; supports choice / noul / score. Latency claim: ~0.3 s without thinking, ~3.3 s median with it (H100); CUDA required.