Open decision model D3 tops Jev Index, ships vLLM recipes from 0.6B to 27B
AccBalanced · x · 2026-10-19
The vLLM Semantic Router team released Decision 3.0 (D3), a family of open multimodal decision models spanning 0.6B to 27B, claiming #1 on both Jev Decision Index 0.3.2 boards (66.3 text, 71.3 vision).
Unlike generative LLMs, D3 returns choices, yes/no answers, and scores with probabilities over text, JSON, images, and video — aimed at routing, classification, and scoring without generating long responses.
- All six models have merged vLLM recipes: vllm serve vllm-sr/d3
- Also servable via the System One API (/v1/systemone) with control over data residency, latency budgets, and governance
- D3 27B delivers 1.8× requests/sec of its checkpoint server on a single MI325X at 64 concurrent clients
More from Infra
- TensorSharp adds a local LLM editing assistant driven by the same commands as its UI — fuzhongkai · 2026-10-19
- Microsoft open-sources LiteBox, a gVisor-like secure library OS for Linux and Windows — mgill25 · 2026-10-19
- Extropic's Beff Jezos Hints TSUs May Soon Power Local Inference — beffjezos · 2026-10-19
- TensorSharp adds a local video studio: LTX 2.5 image-to-video, timeline editing, MP4 export — fuzhongkai · 2026-10-19
- Fiber-based co-packaged optics demo shows few dies in 2025, Intel's COUPE urged to hurry — jwt0625 · 2026-10-19
- LTX 2.5 runs locally on M5 Pro MacBook: 6s video in 13 minutes, fully open-source — fuzhongkai · 2026-10-19