Liquid AI's Decision Models: d1-ultra Tops AA Index, d1 Series Open-Sourced
Liquid AI 的 Leonie Monigatti 发布长推文串，总结公司在「决策模型」(decision models) 这一快速兴起赛道的布局与技术方案。据她所说，自该赛道开启以来 Liquid 已发布 7 款决策模型——3 款 API 托管、4 款开放端侧运行，其中 d1-ultra 目前在 Artificial Analysis 决策指数登顶；开源的 d1 系列已上线 Hugging Face，d1-2B 是专做分类决策的小模型。
已确认
- 开源生态的决策模型分成 encoder 路线(如 convaiinnovations/laya、fastino/GLiNER2.5-De…) 与 decoder 路线两大阵营，Liquid 表示两边都试过，已拿到能与原版 Jev 竞争的配方
- 新版本同时推进三个方向：适配端侧设备、增加更多模态、增加更多决策原语(如带置信度校准的类型化答案输出)
- 核心架构设计：保留 decoder 架构但解除因果掩码。作者推理是因果掩码对「预测下一个 token」的语言模型训练合理，但决策模型可以看到全部 token,解除后获得双向注意力
- 置信度校准：模型刚训练完往往过度自信，自称 90% 确信的判断实际正确率仅约 68%；校准后号称 90% 实际正确率显著贴近(附前后对比图)
- 校准方法上，据说 Typesafe 用 RLCD,但作者指出实践中不需要强化学习，多数团队用单参数的 temperature scaling 即可：T<1 让 token 分布更尖锐，T>1 则更平滑
- 行业动向：据配图，OpenAI Developers 宣布 Decisions API 公测(号称比 GPT 快最多 10 倍)，Cloudflare、微软等也纷纷入局
为什么重要
- 置信分数正是这类决策模型有用性的关键，作者认为决策模型的用例与可叠加原语仍处早期探索阶段，几周内赛道变化极快
- d1 系列开源且可端侧运行，为开发者提供了低成本试用这一新范式的入口
2026-11-06 ~ 2026-11-06 · 9 related posts
Primary sources
- Liquid AI says d1-ultra tops Artificial Analysis Decision Index with 7 decision models released — helloiamleonie ·
- Liquid AI lifts the decoder's causal mask: bidirectional attention beats causal for decision models — helloiamleonie ·
- Liquid AI open-sources d1 decision models: a 2B model purpose-built for classification calls — helloiamleonie ·
- Screenshot collage shows OpenAI, Cloudflare, Microsoft all racing into decision models — helloiamleonie · 2026-11-06
- [source] Liquid AI says d1-ultra tops Artificial Analysis Decision Index with 7 decision models released — helloiamleonie · 2026-11-06
- Open decision models split into encoder and decoder camps — Liquid AI tried both — helloiamleonie · 2026-11-06
- Liquid AI lifts the causal mask: decoders can be decision models too — helloiamleonie · 2026-11-06
- [source] Liquid AI lifts the decoder's causal mask: bidirectional attention beats causal for decision models — helloiamleonie · 2026-11-06
- Freshly trained decision models are overconfident: claims 90% sure but only 68% correct — helloiamleonie · 2026-11-06
- No RL needed: temperature scaling is the simple trick teams use to calibrate model confidence — helloiamleonie · 2026-11-06
- Liquid AI's decision recipe matches Jev; new release adds on-device, modalities, primitives — helloiamleonie · 2026-11-06
- [source] Liquid AI open-sources d1 decision models: a 2B model purpose-built for classification calls — helloiamleonie · 2026-11-06