Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options, Not Text

Nace.AI has put Drex 1.5 out in the open, and the unusual part is what it refuses to do. This is not a chatbot and it won’t draft a tidy paragraph for you. It looks at a state, reads typed questions, and scores the options you give it. No sampling. No temperature. Just probabilities across a fixed set of answers.

The model sits at 8.95B parameters and comes with bf16 weights of about 18 GB. By default it uses a 16,384-token context, but Nace says it can stretch to 131,072. On the hardware side, that means one CUDA GPU in bf16, with testing done on a 24 GB A10G. There’s also a Q8_0 GGUF at about 9.5 GB that runs on Apple silicon and CPU.

Performance is the pitch here, and Nace has the numbers to make the pitch awkward for the closed camp. On the public Decision Index 0.3.1, Drex 1.5 scores 58.08. That edges into the board’s tie band with Jev 1.13.0, which scores 57.96, and Nace says Drex leads Jev on 20 of 37 benchmarks. The strongest area score is Tools at 75.0. Knowledge and Reasoning sit much lower, at 44.6.

Long documents look like Drex’s sweet spot. Nace says it hits 93.4% accuracy on 32K to 128K-token documents, with a median 2.0 seconds, while truncating those same requests to 8K drops accuracy to 76.5% and 78%. But the model also has obvious blind spots: 45.4% on GPQA Diamond, 58.7% on MMLU-Pro, and just 7.4% per-review F1 on ACOS aspect sentiment. So yes, it’s a strong specialist. It is not pretending to be a general-purpose brain in a box.

Developers can run it through Python, llama.cpp, Ollama, or a hosted OpenRouter version priced at $0.04 per 1M input tokens and $0 output. Nace also says existing Jev clients can work with only a few environment variable changes. That is the real headline: a closed-category workflow model just became something teams can download, run locally, and poke at without asking permission.