Drex

Drex is the decision model from Nace.AI, a Palo Alto company founded in 2024 that also sells document-processing models to finance and audit teams. It is a System One model. You send a state and typed questions, and one forward pass per question returns a probability for every option. It writes no text. Nace launched Drex 1.0 on 24 September 2026 and moved its drex-latest alias to Drex 1.5 on 28 September. Drex 1.5 is a 9B model post-trained from Xiaomi’s MiMo-V2.6-Distill-Qwen-9B, with a pointer head that scores each option. It reads states of up to 131,072 tokens, where Drex 1.0 read 32,768.

The API is POST https://console.nace.ai/v1/systemone, the same request and response as Jev. Nace’s migration guide says TypeSafe’s SDK works with a changed base URL, key and model name, though Drex rejects a few loose inputs Jev accepts, such as non-string instructions. A Choice returns the chosen option and a probability per option, a Noul the probability of yes, and a Score a probability-weighted level. A request holds up to 512 questions.

Every number here is vendor-run. Nace scored Drex 1.5 itself with the official Decision Index 0.2.1 kit on 28 September 2026: 58.28 against 57.91 for Jev 1.13.0, ahead on 21 of 38 benchmarks. Nace says Drex trained on those benchmarks’ training splits and was scored on their held-out splits. On 231 public JevBench items it got 199 right to Jev’s 201. In its own board-game arena it won 122 games to Jev’s 87, with 47 draws. Drex is not on the public Decision Index 0.3 or Benchmark Heaven’s JevBench as of 10 October 2026.

Nace aims it at the decision between agent steps: which document to read next, which tool to call, whether to escalate. That fits support inbox triage, typed tool dispatch, confidence-gated actions and agent routing. Its long context suits whole contracts or case files.

By Nace’s own figures it trails Jev on graduate-level science (GPQA Diamond, 45.4% against 78.6%), broad knowledge and aspect-level sentiment. It reads text only.

Drex is generally available on Nace’s API, with $25 of free credit at signup, then $0.05 per million input tokens. OpenRouter serves it through DeepInfra at $0.04. The 1.5 weights are on Hugging Face, but the Nace.AI Open RAIL-M licence bars use beyond personal or research work by any organisation with more than $1 million in yearly revenue or funding raised, and any use by a competitor of Nace, without a separate licence. Nace also publishes Drex DLM, a research model on NVIDIA’s Efficient-DLM-8B under CC BY-NC 4.0.

Question types Choice Score Noul Max Choice optionsNot documented Score levelsNot documented Questions per call512 Total context139,264 tokens State budget131,072 tokens Rate limit120 requests per minute and 8 in flight on the free tier; 600 requests per minute and 16 in flight on the paid tier, which an account joins with its first top-up. Nace can set custom limits per account. EndpointPOST https://console.nace.ai/v1/systemone SDKsTypeScript: nace-sdkPython: nace-sdk drex-v1.5 takes a state of up to 131,072 tokens, and the state plus the longest question may be up to 139,264 tokens. drex-v1.0 allows 32,768 for both. The JSON body of state and questions is capped at 1,048,576 bytes, and a request holds 1 to 512 questions. States over 8,192 tokens go to a separate long-request pool. Nace waits up to 55 seconds for the model before returning a 529. The self-hosted weights default to 16,384 tokens and can be set up to 131,072. OpenRouter lists a 131,072-token context, checked 2026-10-10. Nace publishes no cap on options per Choice or levels per Score.

drex-v1.5, 28 Sep 2026, The current version and the target of the drex-latest alias since 2026-09-28. About 9B parameters (8,953,803,264 per Hugging Face), post-trained from XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B with a pointer head that scores each option. Weights on Hugging Face as nace-ai/drex-v1.5 under the Nace.AI Open RAIL-M licence, with a Q8_0 GGUF; runs on Python, and on Nace's llama.cpp and Ollama forks. Release notes drex-v1.1, Retired on 2026-09-28. drex-latest pointed to it from 2026-09-27; requests that name it are now answered by drex-v1.5 at its price. Release notes drex-v1.0, 24 Sep 2026, The launch version, still served at $0.04 per million input tokens with a 32,768-token limit. The launch post gives it under 6B parameters. No weights published. Release notes

What people use Drex for, one page per pattern. Workflow controlStarter

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. Choice Score Noul Workflow controlAdvanced

Ask Jev one Choice question over your tool names and one per closed-set argument, so every value that comes back is a value the function already accepts. Nouls decide whether an optional argument was mentioned at all. Your code assembles the call and runs it. Choice Noul Workflow controlIntermediate

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. Choice Score Real-time and agentsIntermediate