Open source, self-hosted. There is no hosted API. You download the weights and run them on your own hardware. How open-source models are listed.
What Torchcast Decision 27B is
Torchcast Decision 27B is a System One model from Torchcast AI. The company’s site says it builds forecasting systems and lists Singapore as its base. The model reads a state, which can be text or JSON, and answers typed questions: Choice, Noul and Score. Each answer carries a probability for every option, and all the questions in one request are answered together, usually in a single pass. Its server speaks the same /v1/systemone request format as Jev, so a client written for Jev should work against it.
How it is built
The model card calls it a LoRA fine-tune of StartLux-Decision-27B, merged into the weights. A LoRA is a small set of trained adjustments added to a larger model. The Jev Decision Index, multimodalart’s community board on Hugging Face, lists the base as Qwen/Qwen3.8-27B. Torchcast trained on the train and dev splits of public datasets, plus decision data built from them. It says no Index test rows were used, and that about 1% of training rows share a source document or question template with a test item. The repository ships a serving package that needs a Linux machine with an NVIDIA GPU of 64 GB or more.
Licence and access
The weights are CC BY-NC 4.0, inherited from StartLux-Decision-27B. That allows non-commercial use only. Commercial use needs separate permission from Torchcast AI and from StartLux Labs, and the training data may add its own limits. Torchcast says the model is for research and non-commercial use. We found no hosted API, and OpenRouter did not list it on 12 October, so it cannot stand in for a hosted model in a paid product.
What the benchmarks show
On the Jev Decision Index, in the file generated on 12 October 2026, it scores 59.91 on the Full score against Jev’s 60.11, and 65.1 on the public score against 57.96. Only one entry has a higher public score. Its expected calibration error, the average gap between stated confidence and actual accuracy, is 0.0235 against Jev’s 0.074. It scores higher on the public tests than on the private ones that make up 80% of the Full score. The row also records that Index results guided Torchcast’s choice of checkpoint, and the card says the same, so the public number is not a clean holdout.
Torchcast’s own run on the older 0.2.1 version gave 65.00, which is vendor-run. Its card says the Knowledge area trails Jev, and Language is slightly below the base model. The 27B model has no row on JevBench.
What it is not for
The card advises against safety-critical, security, medical, legal, financial and employment decisions without human review, and warns that probabilities can be miscalibrated on new domains.
Specifications
| System One Score | 95.5 from 1 of 3 boardsJev Decision Index 95.5. How the score works |
| Question types | ChoiceScoreNoul |
| Max Choice options | Not documented |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | 262,144 tokens |
| State budget | Not documented |
| Rate limit | None. It runs on your own hardware. |
| Endpoint | http://localhost:8090/v1/systemone |
| Licence | CC-BY-NC-4.0 |
| SDKs |
The model card gives a context of up to 262,144 tokens and says a very long state is read once in chunks first. A choice with more than 26 options takes extra rounds. Torchcast's evaluations are text-only and English-only; image input is inherited from the base model but was not evaluated. It needs a Linux machine with an NVIDIA GPU of at least 64 GB, and the bf16 weights are 54.7 GB.
Versions
- torchcast-ai/torchcast-decision-27b, 3 Oct 2026, A LoRA fine-tune of StartLux-Decision-27B, merged into the weights. The Jev Decision Index lists the base as Qwen/Qwen3.8-27B and 27.8B served parameters. The Hugging Face account also has 12B and 4B checkpoints. Release notes
Use cases
What people use Torchcast Decision 27B for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
Confidence-gated actions with System One models
Jev returns a confidence value from 0 to 1 alongside every Choice and Score answer. Your code treats it as a separate axis: act automatically when it's high, confirm or flag when it's middling, hand the decision to a person when it's low. Riskier actions get higher bars.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Jev Decision Index 0.3.1 | Full score | 59.91 score | Independent | 12 Oct 2026 |
| Jev Decision Index 0.3.1 | Public score | 65.1 score | Independent | 12 Oct 2026 |
| Jev Decision Index 0.3.1 | Expected calibration error | 0.0235 ECE | Independent | 12 Oct 2026 |
| Jev Decision Index (Torchcast report) 0.2.1 | Overall skill score | 65 score | vendor-run | 3 Oct 2026 |
- Jev Decision Index, Full score: Row: Torchcast Decision 27B. Counts 20% public benchmarks and 80% private tests; 0 is random guessing, 100 is perfect. The row notes that Index results guided checkpoint selection (author-declared). Jev scores 60.11.
- Jev Decision Index, Public score: Row: Torchcast Decision 27B. Second-highest public score in the file. Jev scores 57.96.
- Jev Decision Index, Expected calibration error: Row: Torchcast Decision 27B. Lower is better. Jev scores 0.074.
- Jev Decision Index (Torchcast report), Overall skill score: Torchcast ran the official 0.2.1 kit itself on all 155,390 suite rows; its card lists Jev 1.13 at 57.91. The card says Index results guided checkpoint selection, so it is not an untouched holdout.