Best AI model under 32B for decision models — October 2026
As of Oct 14, 2026, vLLM-SR d3 (27B) & linne-oak 1.0 (31B) & Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout) are tied for state of the art with low confidence in Decision models, models at or under 32B (since Oct 14, 2026; sources Decision Index Full and JevBench; methodology v1.0).
Category board for Decision models
Ranks
- #1 vLLM-SR d3 (27B)
- #1 linne-oak 1.0 (31B)
- #1 Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout)
- #2 Sieve 31B
- #3 Bobcat 1.3 (31B)
- #4 decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted)
- #5 torchcast-decision-27b (Torchcast AI, Qwen3.8-27B fine-tune)
- #6 Darwin-27B-ZTC
- #7 Sieve 27B
- #8 René-1 31B FP8 (salfatigroup, Gemma 4 31B full fine-tune, one-pass option readout)
Last crown changes
- Oct 14, 2026 — Decision models: vLLM-SR d3 (27B) and linne-oak 1.0 (31B) and Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout) are co-SOTA
- Oct 8, 2026 — Decision models: Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout) takes the crown
Which benchmarks count
- Decision Index Full — primary, weight 1
- JevBench — primary, weight 1
- Decision Index latency — display, weight 0
- JevBench price — display, weight 0
- Vendor decision claims — display, weight 0
Who is SOTA in Decision models, models at or under 32B?
As of Oct 14, 2026, vLLM-SR d3 (27B) & linne-oak 1.0 (31B) & Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout) are tied for state of the art with low confidence in Decision models, models at or under 32B (since Oct 14, 2026; sources Decision Index Full and JevBench; methodology v1.0).
Since when is the Decision models, models at or under 32B crown current?
The current crown in Decision models, models at or under 32B has held since Oct 14, 2026.
What is a decision model?
A decision model takes state, as text or as images, plus a typed question. It returns a typed answer with probabilities rather than free text. Output types: noul (yes/no/unknown), choice, score, multi-label.
How is calibration ranked?
The calibration slice ranks the lowest ECE among models in the top quartile of accuracy, so a constant probability cannot take the slice. Brier breaks ECE ties.
Do vendor-reported decision numbers count?
Vendor-reported numbers are display-only at weight 0.