Lumma-Fev-27B 决策模型发布，准确率超越 TypeSafe Jev
原标题：Introducing Lumma-Fev-27B
AI 摘要
Hugging Face Blog 发布 Lumma-Fev-27B，这是 Lumma-Fev 决策模型家族中最大的新成员，基于 Qwen3.8-27B 全权重训练约 10 万条决策样本，支持英语及印地语、泰米尔语、卡纳达语。在 4,454 条留出决策上准确率达 90.7%，高于 TypeSafe Jev 的 83.9%，校准误差也更低（0.009 对 0.045），但在 BBH、WinoGrande 等推理基准上仍落后于 Jev。模型以 Apache 2.0 许可开源，沿用与 Jev 相同的 /v1/systemone 接口。
正文节选
In September we released Lumma-Fev, a family of decision models that score the options you supply in a single forward pass instead of generating a label for you to parse. The family ran from 154M to 9B parameters. Today we are adding the biggest one. Lumma-Fev-27B starts from Qwen3.8-27B and trains every weight, with no LoRA and no frozen layers, on about 100K decision samples. It works in English and three Indic languages: Hindi, Tamil and Kannada. It is the first Fev model we compare head-to-h