vLLM-SR d3 (27B) vs Decider 31B (NVFP4)
vLLM-SR d3 (27B) is more accurate on the Jev Decision Index, Decider 31B (NVFP4) is better calibrated. Choose on whichever of accuracy, calibration and cost matters most for your decisions.
| Spec | vLLM-SR d3 (27B) | Decider 31B (NVFP4) |
|---|---|---|
| Provider | vLLM Semantic Router | Mapika |
| Decision Index score | 66.3 | 64.6 |
| Calibration error (ECE) | 0.043 | 0.024 |
| Median latency (p50) | 97 ms | 72 ms |
| Cost per 1,000 decisions | — | — |
| Parameters | 26B | 33B |
| Open weights | Yes | Yes |
| Question types | Choice (one of N options), Score (ordered rubric), Yes / no | Choice (one of N options), Score (ordered rubric), Yes / no |
| Mechanism | Decoder LM with a decision readout | Decoder LM with a decision readout |
| License | apache-2.0 | apache-2.0 |
FAQ
Is vLLM-SR d3 (27B) or Decider 31B (NVFP4) more accurate?
vLLM-SR d3 (27B) scores higher on the Jev Decision Index (66.3 vs 64.6).
Which is better calibrated, vLLM-SR d3 (27B) or Decider 31B (NVFP4)?
Decider 31B (NVFP4) has the lower expected calibration error (0.024 vs 0.043), so its stated confidence is closer to how often it is right.
Is vLLM-SR d3 (27B) or Decider 31B (NVFP4) cheaper per 1,000 decisions?
A hosted price is not recorded for both models; open-weight models cost the compute you run them on.
Cost per 1,000 decisions assumes 200 input tokens per decision. Pricing is indicative — confirm with the provider before production use. Highlighted values indicate the better figure for that row.