Open Decision 3.0 Models From 0.6B to 27B Claim #1 on Both Jev Decision Index Boards
henzy123 · reddit · 2026-10-12
The vLLM Semantic Router team released Decision 3.0 (d3), an Apache-2.0 open-weights decision model family spanning 0.6B to 27B, claiming #1 on both Jev Decision Index boards.
Technical details
- Built on Qwen3.5 (0.8B/2B/4B/9B) and Qwen3.8-27B; parameter counts include vision encoders;
- Three decision types — choice, yes/no, score — emitting probabilities with no text generation; multiple questions per input answered in one call;
- Multimodal: multiple images per request (up to 1.6MP) and videos (2fps sampling, max 32 frames);
- BF16 weights, Transformers support, trained on AMD Instinct MI325X.
Results (official Jev 0.3 kit)
- 27B scores 64.7 text / 71.6 vision, ahead of every listed model including Jev itself;
- Every size leads its class; the 9B d3-flash beats the previous-gen 27B;
- Video zero-shot (no video in training): 27B hits 77.4 and 4B 70.3 on all 19,140 Perception Test questions.
Related event: Open-Source Multimodal Decision Model Decision 3.0 Released(2 posts)→
More from Models
- Did GPT-4.5 understand music better than any human savant? Omni-modal waveform training debated — RileyRalmuto · 2026-10-12
- Google testing new Flash-Lite checkpoints as Gemini Pro tier stays frozen for 235 days — lyraxana · 2026-10-12
- Users Complain Gemini Web Locked Into Auto Mode With Only Flash-Lite — Slow_Pangolin_5494 · 2026-10-12
- Haiku 5.5 costs 75% less than 4.5, but shines only as a subagent under Opus 5.5 — Scobleizer · 2026-10-12
- Power user burns $6,428 in API-priced tokens on two $200 plans: 'VC-subsidized scooter era' — beffjezos · 2026-10-12
- A 'continuity recipe' for making LLMs keep a persistent identity goes viral — RileyRalmuto · 2026-10-12