Give Rune text, data or an image, a question and your options. It answers with one of them and a calibrated probability for every one, in a single pass. Open weights, so it runs on your own cards.
Rune reads the input, answers your question with one of your options, and tells you how sure it is. A bank complaint, an insurance claim, a phishing email, a contract clause, an invoice, a chart: the same model, the same kind of answer.
I was charged twice for the same purchase on 14 March and the second charge still hasn't been reversed.
QuestionWhich team should handle this complaint?
Cards
Real answers from Rune, one request each, recorded 2026-09-28.
Each question names its type, the instruction and the options. The answer is always one of your options, never free text.
Route a complaint, a ticket or a referral to the team that should handle it.
I was charged twice for the same purchase on 14 March and the second charge has still not been reversed.
Which team should handle this complaint?
Cards
One of your three teams, with every probability
Check a fact: does the invoice match the purchase order, does the reply answer the question.
Purchase order: Nordkraft AS, 3 lines, €12,400. Invoice: Nordkraft AS, 3 lines, €12,400.
Does the invoice match the purchase order?
True: it matches
One probability: how likely the statement is true
Severity, risk or urgency, on levels you define, read back as a position on them.
Payments have been failing for all customers in the Nordics since 09:12. Retries are not succeeding.
How severe is this incident?
Down for everyone
1.94 on a 0–2 scale
Ask all three about the same input at once: one request, one pass, three answers.
A choice returns one of the keys you supplied, a score a position on your list. Nothing to parse, nothing invented.
Each answer carries the probability of every option you offered, so a close call and a clear one look different.
Higher confidence means higher accuracy, so apply the sure answers automatically and send the rest for review.
Ask about category, priority and completeness together; each answer stands on its own.
The input can be a string, a record, a list or an image. Send what you already have.
With thinking on, a question Rune is less than 70% sure of reasons briefly first. The rest stay one pass.
A probability is only useful if it is honest. Rune's is: when it says 90%, it is right about nine times in ten, and when it says 60%, about six.
Ten answers Rune gave at 90% sure
About 9 of 10 turn out right.
Ten answers Rune gave at 60% sure
About 6 of 10 turn out right.
So draw a line. Above it, apply Rune's answer automatically. Below it, send the case to a person. You know in advance how many will need a human.
Measured on 33 Decision Index benchmarks with a right answer per field: at the recommended setting, Rune's average confidence is 73.6% and its accuracy 73.9%. The tick rows above illustrate the rule; they are not a recorded run.
The Decision Index scores decision models on 38 benchmarks in five areas, rescaled so that guessing scores zero. Its maintainers reran our submission of the open weights and matched it bit for bit.
Decision Index 0.2.1, #2 on the board as published on 26 September 2026.
Skill per area, 0 = guessing, 100 = perfect.
Scanned forms, photos of meters, screenshots attached to a bug report, charts and tables. Send the picture, and ask what you would ask about the text.
Send each complaint, ticket or referral to the team that should handle it.
Does the invoice match the order? Does the reply answer the question? Is the document still valid?
Incident severity, transaction risk and ticket urgency on a scale you define.
Choose the next tool, step or sub-agent fast enough to sit in the loop itself.
Tag a whole archive with several attributes per document, paying for GPU time and nothing else.
Is this claim consistent with the policy? Is this message phishing? Borderline cases go to a person.
Score a generated reply for accuracy, tone or policy before it reaches a customer.
Pull the supplier, product line or department when the allowed values are known.
In banking, healthcare, insurance and other regulated work, sending data to someone else's API can rule a model out before accuracy comes up. Rune runs inside your environment, offline if you need it.
Send decisions to our endpoint and start without setting up a GPU.
Run it on your own cards and the data never leaves your environment. It works fully offline.
per decision on one NVIDIA RTX PRO 6000, served by Surogate
hf auth login # accept the terms on the model page first hf download surogate/rune-26b-a4b-GGUF --local-dir rune surogate serve ./rune --host 0.0.0.0 --port 8000 \ --max-model-len 32768 --max-num-seqs 8 --kv-capacity auto --decision-temperature 2
Fastest on NVIDIA Blackwell (RTX 50 and RTX PRO). Also runs on Hopper, Ada Lovelace and Ampere. The weights take 51.6 GB in bf16. --decision-temperature 2 makes confidence match accuracy; it never changes which option Rune picks.
Cameras, robot arms, invoices and claims, with the answers Rune really gave.