Introducing Drex.
A small model that decides.
Facts in, a probability for every option out. One pass, and no text to parse.
- 58.28
- Decision Index 0.2.1, rank 1
- 1
- accelerator
- 128K
- tokens of context
- <10B
- total parameters
The axis starts at 55, not zero. Total parameters under each name; Jev's size is not published.
Trusted by
Big Fouraccounting firm
Where Drex fits in an agent. The decision between the steps.
An agent makes many decisions before it produces an answer: which document to read, which tool to call, whether it has enough evidence or should keep searching. They set the speed, cost and quality of the whole workflow. Drex is built for exactly these moments.
Take an agent investigating an invoice discrepancy.
1 · Your search system
Retrieve
Retrieves candidate documents: invoices, contracts, amendments and payment records.
2 · Drex
Evaluate
Scores which documents, or which next actions, are most relevant to the case.
3 · Your code
Execute
Reads the selected files, applies your rules, and runs the next step.
4 · A generative model
Explain
Writes the final explanation once the evidence is in.
The search system retrieves. Drex evaluates the choices. Your code executes the next step. A generative model writes the explanation.
What it receives
The current context, plus a typed question with the options you allow.
state: "Invoice 4471 billed 12,400, contract allows 11,000. Amendment 2 is on file. No payment records read." question: next_step inspect_amendment query_payment_records proceed_to_answer
What it returns
One probability for every option you supplied. Nothing else comes back.
Illustrative values showing the shape of a response.
How your application uses it
The probabilities are ordinary data, ready for your code.
- Rank candidates and read the top few.
- Take the action when one option clearly leads.
- Send the case to a person when the probabilities are close.
The same pattern applies to routing support requests, prioritizing claims, selecting tools, and identifying cases that need human review.
Ask it a question. It tells you how likely each answer is.
Send Drex the facts of a case and the options you would accept. One pass through the model returns a probability for each one. It never writes text, so it cannot answer with an option you did not offer.
You send
Invoice INV-88214 bills 1,200 units at $14.10. PO-5512 authorised 1,200 at $13.90. Goods receipt confirms 1,200 delivered, 2 days late.
What should happen to this invoice?
hold for buyer · approve · dispute
You get back
Example values; recorded answers are further down the page.
It is not a chatbot. It is the other kind of thinking.
The quick, practised call an expert makes on sight, not the essay they would write to justify it.
How it answers
A chat model: Thinks out loud, one word at a time, then names an answer at the end.
Drex: Reads the case once and scores every option in the same pass.
What comes back
A chat model: A sentence, or JSON you parse and hope names one of your options.
Drex: A probability for each option you offered. Nothing else can come back.
A distribution, not a verdict.
Because the answer is a probability for every option, you can see how close the call was. Threshold it, route the uncertain cases to a person, and act on the confident ones without reading a word.
Because Drex cannot generate text, it cannot invent an option outside your set. That does not make the chosen option correct: the context, the choices and your rules for acting still matter, and your code stays responsible for permissions and execution.
Each board is an example case, and each pile is the probability on one option. Recorded answers are further down the page.
First on the public Decision Index. Ahead of every other model.
Decision Index 0.2.1 scores 38 tests of deciding rather than writing: contracts, causes, tool calls, routing, judgment. The index is chance-corrected, so guessing at random scores zero. Drex 1.5 leads all 71 entries with under 10B parameters; every other model in the top ten with a published size has 25.8B to 32.7B.
| Drex 1.5 | Jev 1.13.0 | Surogate Rune v3 | Decider chat · Gemma | AutoJev-27B | |
|---|---|---|---|---|---|
| OverallDecision Index 0.2.1, 38 tests, chance-corrected. Rows below show each test's own metric. | 58.28 | 57.91 | 57.44 | 57.33 | 56.40 |
| Knowledge & ReasoningCLadder · accuracy | 86.2% | 72.6% | 72.7% | 74.6% | 74.5% |
| Knowledge & ReasoningBBH · accuracy | 85.7% | 92.9% | 81.2% | 76.2% | 78.1% |
| Language UnderstandingContractNLI · macro-F1 | 82.9% | 71.7% | 77.0% | 73.6% | 78.1% |
| Language UnderstandingACOS · per-review F1 | 7.4% | 29.5% | 26.8% | 18.8% | 20.4% |
| Retrieval & ClassificationAmazon ESCI · macro-F1 | 56.2% | 55.2% | 55.3% | 53.1% | 55.1% |
| Knowledge & ReasoningGPQA Diamond · accuracy | 45.4% | 78.6% | 46.9% | 49.0% | 49.5% |
| Retrieval & ClassificationBRIGHT · nDCG@10 | 44.3% | 47.5% | 46.3% | 43.6% | 48.7% |
| Tools & AutomationToolRet · nDCG@10 | 70.3% | 65.3% | 64.4% | 63.6% | 67.4% |
| Tools & AutomationAPI-Bank · accuracy | 85.2% | 88.2% | 83.3% | 85.0% | 84.1% |
| Arts & Human TasteHabermas Machine · accuracy | 59.5% | 45.9% | 42.2% | 47.9% | 41.8% |
| Arts & Human TasteBPoMP · accuracy | 83.3% | 90.9% | 90.0% | 90.6% | 93.9% |
| Knowledge & ReasoningCRUXEval · accuracy | 75.3% | 73.0% | 74.4% | 79.3% | 74.9% |
| Knowledge & ReasoningHLE · accuracy | 11.8% | 20.4% | 10.6% | 14.8% | 12.0% |
| Language UnderstandingFinEntity · macro-F1 | 87.9% | 87.0% | 88.4% | 90.9% | 92.9% |
| Language UnderstandingANLI · macro-F1 | 70.8% | 74.8% | 73.5% | 73.2% | 70.7% |
| Retrieval & ClassificationBANKING77 · macro-F1 | 85.9% | 79.7% | 83.8% | 79.1% | 79.0% |
| Tools & AutomationBFCL · case exact accuracy | 95.0% | 95.8% | 94.8% | 98.0% | 97.6% |
| Tools & AutomationWhen2Call · accuracy | 89.8% | 81.0% | 76.0% | 76.9% | 81.7% |
| Arts & Human TasteForecastBench · Brier skill vs always 0.5 | 25.7% | 30.6% | 18.9% | 19.2% | 22.1% |
| Arts & Human TasteHumicroedit · accuracy | 63.4% | 61.9% | 62.2% | 63.8% | 62.4% |
| Knowledge & ReasoningMMLU-Pro · accuracy | 58.7% | 82.7% | 67.2% | 69.6% | 64.5% |
| Knowledge & ReasoningChessBench · accuracy | 21.9% | 17.2% | 18.7% | 22.3% | 17.2% |
| Language UnderstandingHellaSwag · accuracy | 92.0% | 94.5% | 92.4% | 92.0% | 94.0% |
| Language UnderstandingRAGTruth · F1 on hallucinated class | 82.8% | 76.5% | 76.8% | 76.9% | 80.3% |
| Tools & AutomationHome appliance simulator · case exact accuracy | 40.9% | 52.3% | 46.6% | 62.5% | 73.9% |
| Retrieval & ClassificationCLINC150+OOS · macro-F1 | 91.6% | 89.3% | 87.4% | 91.2% | 87.8% |
| Arts & Human Tastecfcolor · accuracy | 59.8% | 64.4% | 62.6% | 62.6% | 64.4% |
| Arts & Human TasteNew Yorker · accuracy | 73.3% | 70.1% | 74.1% | 73.9% | 70.3% |
| Knowledge & ReasoningMuSR · accuracy | 57.3% | 66.1% | 64.9% | 64.2% | 61.8% |
| Knowledge & ReasoningGSM8K · accuracy | 83.0% | 79.9% | 79.7% | 82.6% | 67.4% |
| Language UnderstandingNLI4CT · macro-F1 | 81.0% | 84.1% | 80.4% | 82.8% | 84.8% |
| Language UnderstandingVAST · macro-F1 | 76.9% | 64.6% | 76.6% | 73.2% | 70.8% |
| Language UnderstandingWinoGrande · accuracy | 88.1% | 92.0% | 86.0% | 83.7% | 85.2% |
| Retrieval & ClassificationHoVer · accuracy | 86.1% | 72.9% | 80.7% | 76.4% | 74.2% |
| Arts & Human TastePOP909-CL · accuracy | 73.1% | 16.6% | 66.1% | 27.9% | 37.8% |
| Knowledge & ReasoningSATA-Bench · case exact accuracy | 27.0% | 26.4% | 34.8% | 29.3% | 29.9% |
| Language UnderstandingiSarcasmEval · Sarcasm F1 · track A, English | 55.1% | 50.5% | 60.4% | 51.1% | 60.3% |
| Retrieval & ClassificationPhishNChips · accuracy | 67.4% | 62.5% | 80.9% | 87.5% | 60.0% |
Drex 1.5 was scored on 2026-09-28 with the official 0.2.1 kit on all 38 benchmarks. Benchmark cells are each test's own metric; the overall row is the chance-corrected index, where the board treats gaps under 0.25 as ties. Other entries are the public board's published scores, read 2026-09-28.
Join the Drex communitySame kind of model as Jev. 5× fewer tokens per decision.
Jev is the decision model the public index is named after. Drex answers the same requests in the same shape, leads it on the index, and finishes 2 items behind it on Jev's own benchmark.
- Jev 1.13.0
- 87.0% · 201/231
- Drex 1.5
- 86.2% · 199/231
- Drex 1.1
- 82.2% · 190/231
Reads the whole document. Up to 128K tokens.
Drex 1.5 takes up to 128K tokens in one request, up from 32K on Drex 1.1. On long documents with typed questions it is more accurate reading everything than reading the first 8K, and the gap holds past 32K.
Accuracy of Drex 1.5 on long documents with typed questions. Every request was answered, one forward pass per question. "First 8K only" is the same requests cut to their first 8K tokens.
See it answer. Real cases, recorded replies.
Pick a case. The same text and questions went to both columns, and neither wrote a sentence back.
A topic choice and an escalate yes/no on the customer message from Jev's own docs.
Customer: I was charged twice last Tuesday and I am furious. Order #4419. I want a refund today.
topic choice
What is the issue about?
billing · bug · shipping
urgent yes/no
Escalate to a human now?
Drex
72 tokens read
topic choice
urgent yes/no
Jev 1.13.0
365 tokens read
topic choice
urgent yes/no
Run your own case.
Paste a case, name the options, and get the probabilities back. Free to try, no card needed.
On topic, both columns are all but certain; the other questions show how far apart they can be.
Runs where your data lives.
Small enough to own, and to keep. Every way of running Drex answers the same request in the same shape, so you can mix them.
Managed
On our GPUs, shared or dedicated, with capacity, scaling and upgrades handled for you.
Your Cloud
Inside your AWS or Azure account. The model fits on one accelerator, so decisions on regulated data never leave your network.
Air-Gapped
Your datacenter, your hardware, no outbound connection. Updates arrive as signed offline bundles.
Tuned on your decisions
The results on this page come from training on decision data. We tune Drex on your labelled outcomes and ship you the weights.
Drex 1.1 in FP8 decides like it does in bf16.
The FP8 weights take about half the memory. On 2,511 questions they gave the same answer as bf16 on 98.1%, moved probabilities by 0.015 on average and 0.29 at most, and scored 92.57% accuracy against 92.49% in bf16.
SOC 2 Type II. Audited, not promised.
Independent auditors test the controls around Drex, so it can go where the decisions are regulated.
Security controls
Our Type II report covers how Drex is built, deployed and supported, with the security controls tested over a period of operation.
Penetration test
The report from a penetration test of our environment, available on the trust center.
Shorter vendor reviews
Send security and procurement the report instead of filling in another questionnaire.
Data stays where you put it
Run it self-hosted or hybrid and the cases it reads never leave your network.
Every decision on the record
Each answer is a probability over the options you supplied, so every decision can be logged and reviewed.
Drex 1.5 also beats Jev at board games, 122 to 87.
Eight games, 256 matches, every legal move offered as an option. Neither model looks ahead; each move is one pass over the board. Drex 1.5 won more matches than Jev in five of the eight, and in the total.
- Lines of Action26–3–385.9%65.6%
- Checkers21–3–870.3%59.4%
- Chess9–21–260.9%57.8%
- Dots and Boxes18–0–1456.3%53.1%
- Clobber17–0–1553.1%50.0%
- Reversi15–1–1648.4%37.5%
- Nine Men's Morris5–19–845.3%43.8%
- Four in a Row11–0–2134.4%68.8%
A head-to-head we ran between Drex 1.5 and Jev 1.13.0 with the Kaggle Game Arena harness, OpenSpiel rules. Every game has 16 openings, each played from both sides, so neither model keeps the first move. Score counts a win as one and a draw as a half. The last column is Drex 1.1's score against Jev under the same protocol.
Questions, answered.
What is a decision model, and how is it different from an LLM?
A decision model answers closed questions with numbers instead of writing text. You send the facts and the options; one forward pass returns a probability for every option. There is no token stream, no chain of thought and no JSON to repair, so the failure modes of chat models (truncation, malformed output, invented options) do not arise. Drex can only ever answer with options you gave it.
What kinds of questions can it answer?
Three. Pick one of several: named options in, the chosen one plus a probability for each out. Yes or no: the probability that the answer is yes. Rate on a scale: ordered levels in, the expected level plus the spread across levels out. Several questions about one case can share a single pass.
How does it compare to Jev?
Drex 1.5 scores 58.28 on Decision Index 0.2.1, ahead of Jev 1.13.0 at 57.91 and of the strongest community entry, Surogate Rune 26B-A4B v3, at 57.44. It is ahead of Jev on 21 of the 38 benchmarks, with its largest margins on chord recognition, causal inference, group consensus, multi-hop fact checking and stance detection, and it trails on graduate-level science, broad subject knowledge and aspect-level sentiment. On JevBench it gets 199 of 231 items right to Jev's 201. Every result, including every loss, is on this page.
How big is it?
Drex 1.5 has under 10B parameters, the smallest model in the top ten of Decision Index 0.2.1. Every other top-ten model with a published size has 25.8B to 32.7B parameters, and Jev's size is not published. The strongest other model under 10B on the board, JPT-9B, scores 46.89, more than eleven points below Drex. Sizes are total parameters: mixture-of-experts models such as Surogate Rune 26B-A4B use only part of their weights per request, but still have to hold all of them in memory. It reads up to 128K tokens of context in one request.
Is it trained on the benchmarks it is measured on?
It is trained on the official training splits of the index benchmarks alongside verifiable procedural data, and evaluated only on the held-out splits with the official scorer. That is how the leaderboard is designed to be run. It also explains the shape of the results: the model is strong where it has seen the kind of decision before, which is exactly the property that makes it worth tuning on your decisions.
Can it run inside our environment, on our decisions?
Yes. It fits on a single accelerator, so it runs in your cloud account, at the edge or on-premises, as well as on Nace-managed infrastructure or a mix of the two. Every deployment answers the same request in the same shape. We also tune Drex on your labelled outcomes and ship you the weights.
Put a decision model where the guesswork is.
Free to try today. Tuned on your decisions and served in your environment when you are ready.