On October 1, NACE AI released Drex 1.5, a decision model it is pitching as the challenger to Jev. It is the first model of its size to read 128K tokens at once, it took first place on the public Decision Index 0.2.1 ahead of Jev, and you can try the first 250 million tokens for free.
Source: X @NaceAI
Drex 1.5 outputs choice probabilities instead of answers
While a typical language model continues a sentence, a decision model takes the current state, a question, and the possible options, and returns a probability for each one. Drex handles choice, yes/no, and rate decisions in a single inference pass. Rather than generating free text, it returns the judgment values an agent needs to pick its next action. NACE describes Drex 1.5 as a model under 10B parameters.
128K context and long-document evaluation
Drex 1.5 expands input from the 32K of the earlier Drex 1.1 to as much as 128K tokens. In NACE's own long-document evaluation, accuracy was 93.4% when the model read the entire 32K–128K document and 78.0% when it read only the first 8K. Median request latency for that long range is 2.0 seconds.
For the 8K–32K range, reading the full document scored 89.5% and reading only the first 8K scored 76.5%, with a median latency of 0.65 seconds. Documents of 8K–32K get an answer in 0.65 seconds, under a second, and even at 128K it is 2.0 seconds.
It edges out Jev on Decision Index
On Decision Index 0.2.1, Drex 1.5 scores 58.28 and Jev 1.13.0 scores 57.91. They are followed by Rune at 57.44, Decider at 57.33, AutoJev at 56.40, and simple at 55.74. NACE announced that Drex ranks first on this metric.
Source: X @NaceAI
The JevBench result is slightly different. In the tally of 231 public items on NACE's product page, Jev got 201 right (87.0%) and Drex 1.5 got 199 right (86.2%), a difference of two questions. In 256 games across eight board games listed on the same page, Drex went 122 wins, 87 losses, and 47 draws against Jev.
It also reads fewer tokens per decision. Using the median of eight identical requests, Drex uses 70 tokens (measured on Drex 1.1) and Jev uses 367.
Pricing and how to try the API
NACE's official announcement lists $0.04 per 1 million input tokens and a free trial of 250 million tokens. NACE says you can try it without a card.
- Enter your email on the Drex login page.
- Log in with the one-time code sent to your email. New users get an account created automatically.
- Issue an API key from the dashboard and use it in the managed environment NACE describes.
Besides the managed service, NACE lists customer VPC, on-premises, and edge as deployment options, and it delivers model weights tuned on customer data. Drex 1.1 is also available with FP8 weights. They use half the memory and gave the same answers as bf16 on 98.1% of 2,511 questions.
NACE AI and the Jev decision model, compared
NACE AI builds small models and agents tailored to business workflows and deploys them on its own cloud, customer clouds, and on-premises. With that approach, Drex enters the same decision-model market as Jev. For how Jev works and its early performance comparison, see the CLM-8B and Jev introduction. For how it is used in development, see jevgrep code search and Pi MCP Codemode and Jev.
Drex 1.5 competes head-on with Jev, leading with long-document judgment at 128K and the top Decision Index score. The decision-model category now has one more option.