D3 六款决策模型已合并 vLLM 配方，27B 版吞吐达 1.8 倍
搞路由和分类的可以看看，D3 六个尺寸都有现成 vLLM 配方，27B 在 MI325X 上吞吐快 1.8 倍，还能跑图片和视频打分。
D3 系列从 0.6B 到 27B 共六款模型，已全部合并 vLLM 配方，可通过 vllm serve vllm-sr/d3 直接启动。在 Jev Decision Index 0.3.2 排行榜上，D3 以 66.3 的文本分和 71.3 的视觉分（10 月 12 日快照）位列两个榜单第一。配方基准测试显示，D3 27B 在单张 MI325X、64 并发客户端的混合负载下，吞吐达到其 checkpoint server 的 1.8 倍。D3 能对文本、JSON、图片和视频返回选择、是/否答案及带概率的评分，适用于路由、分类和打分场景。vLLM 还通过 System One API 提供 D3 全系列、Clef-Flash、Clef-Omni 和 Liquid d1 的推理服务。
Great work from @XunzhuoLiu and the vLLM Semantic Router team on D3. All six models, from 0.6B to 27B, now have merged vLLM recipes!
D3 ranked #1 on both Jev Decision Index 0.3.2 boards: 66.3 text and 71.3 vision (Oct. 12 snapshot).
Run multimodal decisions on your own infrastructure through /v1/systemone—with control over data residency, latency budgets and governance.
vLLM serves decision models too: the full D3 family, Clef-Flash, Clef-Omni and Liquid d1, through the System One API.
D3 returns choices, yes/no answers and scores with probabilities over text, JSON, images and videos. Useful for routing, classification and scoring without generating a long response.
In the recipe benchmark, D3 27B delivers 1.8× the requests/sec of its checkpoint server on 1× MI325X at 64 concurrent clients, using the same mixed workload.
Start with a current vLLM nightly: vllm serve vllm-sr/d3
Recipes for all six sizes: https://t.co/qGDlWJy3bv
API and supported models: https://t.co/L071eXHP7T
