Torchcast Decision 27B
Torchcast AI's decision models: a 27B fine-tune of StartLux-Decision-27B and a 12B fine-tune of Gemma-4-12B-it. Both answer choice, yes/no and score questions with a probability per option through a TypeSafe-compatible /v1/systemone server. Non-commercial weights.
The 27B is a LoRA fine-tune merged into the listed StartLux-Decision-27B, with the MTP head removed. It answers all questions of a typical request in one forward pass, needs a GPU with at least 64 GB, and evaluations are English text only. Torchcast's own run of the Decision Index 0.2.1 suite (a third-party board, official kit, all 155,390 rows) scores 65.00; the board's 0.3 release (6 Oct) places it 4th at 59.91. The 12B (torchcast-ai/torchcast-decision-12b) is a separate Gemma-4-12B-it fine-tune served through vLLM; Torchcast's own run answers 203 of 231 JevBench public items (30 ms p50, 43 ms p95 on one H100), which it calls not an official score. The maker warns that probabilities can be miscalibrated on new domains and that benchmark results guided checkpoint selection. Weights are CC BY-NC 4.0: commercial use needs permission from both StartLux and Torchcast, so they are not mirrored or packaged here. The code is Apache-2.0 (27B) and MIT (12B).
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 27.4B |
| Base model | startlux-models/StartLux-Decision-27B |
| Maker | Torchcast AI |
| Released | 2026-10-03 |
| License | cc-by-nc-4.0 |
| Reported latency | 33.0 ms mean per yes/no question (p95 34.6), 49.0 ms per three-question request (p95 52.1), one H100 (bf16) |
Get the weights
pip install systemonemodels
systemone pull torchcast-ai/torchcast-decision
The files are served from the maker's Hugging Face repository, torchcast-ai/torchcast-decision-27b, and verified against the checksums recorded here.