Veyra Decide
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity4.6/10
Five sizes (0.6B–27B) with full weights on Hugging Face under Apache-2.0, which the company calls temporary while a Veyra licence is prepared. Docs are the cards: Python through trust_remote_code, no server or package. No integrations or SLA.
- Capability6.5/10
Choice / yes-no / score with a probability per option from a decision head on a fine-tuned Qwen3 / Qwen3.5 backbone. Large author tables on public legal and general sets; no latency and no calibration method published. Not on any board.
- Adoption1/100
Repos created 12 Oct: 59 downloads and 2 likes across the five sizes on 14 Oct. No X post, integration or production case found.
Vendor claims
- Veyra Decide 27B: ContractNLI 90.3% vs Jev 77.4%; LegalBench mean of 154 tasks 84.9% vs Jev 82.3%; CUAD clause retrieval mean AUPR 0.651 vs Jev 0.534[Vendor claim — not independently verified]
- Veyra Decide 27B: JevBench 90.0% vs Jev 90.5%; mean of the 13 public Decision Index benchmarks 85.7 vs Jev 86.7; mean of 36 Jev harness tasks 78.4 vs Jev 77.8[Vendor claim — not independently verified]
- Robust to swapped yes/no descriptions: BoolQ 92.7% where the card measures Jev at 9.8%[Vendor claim — not independently verified]
Read this first
- All numbers are JudicialMind’s (UnverifiedClaim). The cards compare against Jev, Clef, Clef-flash and a column labelled “Perplexity” on public sets, measured with inputs up to 16,384 tokens.
- The edge is legal, not general. The 27B leads on contract sets (ContractNLI, LEDGAR, MAUD, LegalBench). On JevBench and the mean of the 13 public Decision Index benchmarks it is just below Jev. That mean is not a Decision Index score.
- The licence may change. “Apache-2.0 for now; a dedicated Veyra licence is in preparation.”
What it is
JudicialMind builds legal agents (intake, research, drafting, negotiation). On 12 October 2026 it published Veyra Decide (Legal & General) on Hugging Face in five sizes: 0.6B, 2B, 4B, 9B and 27B. You pass a document as the state and typed questions (noul, choice, score); model.system_one(state=…, questions=…) returns calibrated probabilities per option.
Each repo ships a full backbone checkpoint plus a separate decision_head.safetensors and a calibration.json. The configs name the backbone: Qwen3 for the 0.6B, Qwen3.5 text for 2B to 27B. The head is a shared bilinear scorer that takes up to 255 options. The 27B card lists 131,072 input tokens.
Author results, 27B (UnverifiedClaim)
| Set | Veyra 27B | Jev | Clef | “Perplexity” |
|---|---|---|---|---|
| ContractNLI (2,091) | 90.3 | 77.4 | 87.8 | 80.7 |
| LegalBench, all 154 tasks | 84.9 | 82.3 | 82.7 | 83.1 |
| CaseHOLD (3,600) | 84.9 | 77.6 | 75.9 | 86.7 |
| JevBench (1,938) | 90.0 | 90.5 | 90.5 | 92.0 |
| 13 public Decision Index benchmarks, mean | 85.7 | 86.7 | 86.4 | 86.0 |
| 36 Jev harness tasks, mean | 78.4 | 77.8 | 77.0 | 80.4 |
The card also tests robustness. With the yes/no descriptions swapped, it measures Jev at 9.8% on BoolQ against 92.7% for Veyra. The smaller sizes trail the 27B: on LegalBench the 9B scores 83.3, the 4B 81.9, the 2B 78.1 and the 0.6B 73.6. No latency is published.
Score working (14 Oct 2026)
- Maturity = availability 8 × 0.30 + docs 6 × 0.25 + integrations 1 × 0.25 + support 2 × 0.20 = 2.40 + 1.50 + 0.25 + 0.40 = 4.55 → 4.6. Availability 8, not 9: the licence is announced as temporary.
- Capability = types 10 × 0.30 + calibration 5 × 0.25 + latency 4 × 0.25 + accuracy 6 × 0.20 = 3.00 + 1.25 + 1.00 + 1.20 = 6.45 → 6.5. Calibration 5: claimed and shipped as a file, method not described. Latency 4: none published. Accuracy 6: many public sets, all author-run.
- Adoption = engagement 2 × 0.35 + ecosystem 0 × 0.35 + production 0 × 0.30 = 0.70 → 1.