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 Rune is
Rune is an open-weights decision model from Surogate, the team behind the open-source Surogate platform for training and serving language models. The Jev Decision Index, multimodalart’s community board on Hugging Face, lists it as a full fine-tune of Google’s Gemma 4 26B-A4B-it, a mixture-of-experts model with 25.8 billion parameters stored in bf16. Surogate’s Hugging Face page describes it as a decision model with image and thinking support.
What it returns
It answers typed questions about a state you supply: Choice, Score and Noul, the same three shapes Jev uses. Surogate’s demo Spaces send a state, a set of questions and optional images to an endpoint, and read back a probability for every option. A Space can ask twelve questions about one draft in a single request. Images go in directly, with no captioning step, which is how the demos read chart screenshots, invoices and damage photos.
What it is good at
On the Jev Decision Index, a zero-shot suite of public and private tests, Rune scores 57.43 overall and 58.28 on the public part, 22nd of 143 rows. On JevBench it holds official rank 20 with 65.0. Its per-type scores favour Choice questions (76.0 public) over Score (56.0) and Noul yes/no questions (42.6).
Microsoft’s launch post for Microsoft-Decision-1 lists Rune at 79.7% accuracy and 91.8 calibration. Microsoft ran that on an undisclosed 36-benchmark suite for its own comparison, so treat it as vendor-run.
What it is not for
Calibration is its weak point. The Jev Decision Index puts its calibration error at 0.1196, with mean confidence of 0.867 against 0.748 accuracy, so a high probability overstates how often it is right. Gate real actions on your own held-out data, not on its raw confidence.
It is also large. The weights come to about 52 GB, so you need a big GPU; the Jev Decision Index and JevBench ran it on a 96 GB card. And the model card is gated, so these facts are unknown as of 12 October 2026: the context window, the option limits per question, Surogate’s own benchmark figures, and any terms beyond the Apache-2.0 tag. Check the card after you accept the gate before you build on it.
Access
Rune is not a hosted API. Hugging Face shows the repo as gated with automatic approval: you accept Surogate’s terms, then download the weights. It had 6,535 downloads and 70 likes on 12 October 2026. No SDK package and no public price were found. Benchmark Heaven says Rune also powers the s1-pro service at system1models.ai, but that site now names a different model for s1-pro, so this page lists no hosted option.
Specifications
| System One Score | 90.1 from 2 of 3 boardsJev Decision Index 91.5, JevBench 88.8. How the score works |
| Question types | ChoiceScoreNoul |
| Max Choice options | Not documented |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | Not documented |
| State budget | Not documented |
| Rate limit | None when self-hosted. Unknown for Surogate's private demo endpoint. |
| Licence | Apache-2.0 |
| SDKs |
Unknown as of 2026-10-12. The model card is gated and was not readable, and no other source states a context window or option limits. The Hugging Face tags list image input. Surogate's demo Spaces ask up to twelve questions in one request.
Versions
- surogate/rune-26b-a4b-GGUF, 21 Sep 2026, Hugging Face repo created 2026-09-21 and last modified 2026-09-26. The Jev Decision Index and JevBench list it as v3. Despite the GGUF in the name, the public file list shows 11 bf16 safetensors shards (25.8 billion parameters) and no .gguf file. Release notes
- invergent-ai/surogate, The Surogate platform repo for training and serving LLMs (Apache-2.0, 854 stars on 2026-10-12). The Jev Decision Index links it as Rune's code. Release notes
Use cases
What people use Surogate Rune 26B-A4B 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.
Compliance verification with System One models
A compliance review is a fixed checklist run against a changing document. Jev answers the whole checklist in one request, returning a probability, a label or a rated level per item, so code can clear the clear-cut findings and send the borderline ones to a person.
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 | 57.43 score | Independent | 12 Oct 2026 |
| Jev Decision Index 0.3.1 | Public score | 58.28 score | Independent | 12 Oct 2026 |
| Jev Decision Index 0.3.1 | Expected calibration error | 0.1196 ECE | Independent | 12 Oct 2026 |
| JevBench 1.7.46 | JevBench score | 65 score | Independent | 12 Oct 2026 |
| JevBench 1.7.46 | Capability Score | 73.5 score | Independent | 12 Oct 2026 |
| Microsoft 36-benchmark comparison | Accuracy | 79.7 % | vendor-run | 10 Oct 2026 |
| Microsoft 36-benchmark comparison | Calibration | 91.8 score | vendor-run | 10 Oct 2026 |
- Jev Decision Index, Full score: Row: Surogate Rune 26B-A4B v3, bf16, 22nd of 143 rows by this score. Counts 20% public benchmarks and 80% private tests; 0 is random guessing, 100 is perfect.
- Jev Decision Index, Public score: Row: Surogate Rune 26B-A4B v3.
- Jev Decision Index, Expected calibration error: Row: Surogate Rune 26B-A4B v3. Lower is better. Mean confidence was 0.867 against 0.748 accuracy, so it runs over-confident.
- JevBench, JevBench score: Official rank 20, 95% interval 62.8 to 66.2. Intelligence 56.1, calibration 90.9, speed 87.0, cost 49.6. The board re-measures rows, so this can differ from earlier editions.
- JevBench, Capability Score: Capability view of the same board. Per-type capability on the board: Choice 76.0 public and 73.1 sealed, Score 56.0 and 54.4, Noul 42.6 and 34.3.
- Microsoft 36-benchmark comparison, Accuracy: Run by Microsoft on an undisclosed 36-benchmark suite (147,137 questions) to market its own model. Mean score across the suite.
- Microsoft 36-benchmark comparison, Calibration: Run by Microsoft on an undisclosed 36-benchmark suite. 100 means confidence matches how often the model is right.