# Torchcast AI releases Torchcast Decision 27B, built on StartLux weights

> A LoRA fine-tune of StartLux-Decision-27B that speaks the /v1/systemone API. Weights are CC BY-NC 4.0; card reports 65.00 on Index 0.2.1.

Published 2026-10-03

Canonical: https://shipped.anth.us/decisions/torchcast-decision-27b/

Torchcast AI published torchcast-decision-27b on Hugging Face on 3 October 2026, by repository creation date. The card describes a 27B model that answers typed questions about a state: multiple choice, yes/no and ordinal rating. It returns a probability for every option and answers all questions in a request in one forward pass.

The base is StartLux-Decision-27B with a LoRA fine-tune merged in. The weights are CC BY-NC 4.0, inherited from that base. The inference code is Apache-2.0. It uses the TypeSafe-compatible /v1/systemone API, so clients written for Jev should work unchanged. Weights are about 54.7 GB in BF16, so it needs an NVIDIA GPU with at least 64 GB of memory. Context is up to 262,144 tokens.

On Decision Index 0.2.1 the card reports Torchcast 65.00, StartLux-Decision-27B 63.88 and Jev 1.13 at 57.91. It says it leads Jev on 31 of 38 index benchmarks and trails it on knowledge-heavy tasks. Latency on one H100 is 33.0 ms for a single yes/no question.

The card lists its limits: single-run estimates, English text only, images not evaluated, and not for high-stakes decisions without human review.

A clean card. The improvement over its parent is 1.12 points.

## Sources
- https://huggingface.co/torchcast-ai/torchcast-decision-27b

## Editor's first pass (AI-assisted)
- Real 3.5: Weights and Apache-2.0 inference code are public; weights are non-commercial.
- Useful 3: Needs a GPU with at least 64 GB; limited by licence for commercial work.
- Notable 2.5: Shows how fast fine-tunes of fine-tunes are stacking.
- Novel 2.5: A fine-tune of another decision model.