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

Nace.AI has open-sourced Drex 1.5, a 9-billion-parameter decision model built to score options rather than write prose. Instead of generating text, it reads a state and a set of typed questions, then returns a probability for every option it is given, a design aimed at agents and backend workflows that must choose between concrete actions.

The release centers on a narrow, deliberate capability. Drex 1.5 cannot generate text, code or explanations; Nace positions it as a decision layer sitting alongside a general model rather than replacing one. According to the company, it scores 58.08 on the public Decision Index 0.3.1, which Nace presents as the top score in the benchmark's sub-10-billion-parameter class.

The published figures are more nuanced than the headline number. Nace notes that Drex 1.5, Jev and Nimble sit inside the leaderboard's 0.9-point tie band, and that Drex leads Jev on 20 of 37 benchmarks. Its area scores are strongest in Tools (75.0) and weakest in Knowledge and Reasoning (44.6). Nace also points out that its own launch chart used the older Decision Index 0.2.1, where Drex scored 58.28 against Jev's 57.91.

Head-to-head comparisons show where the model wins and where it struggles. On JevBench's 231 public items, Drex scores 86.2% against Jev's 87.0%, with both reaching 73.9% on hard items, while across eight OpenSpiel games Drex recorded 122 wins, 47 draws and 87 losses against Jev (56.8%). Knowledge-heavy tests are its clear weak spot: 45.4% on GPQA Diamond versus Jev's 78.6%, and 58.7% on MMLU-Pro versus 82.7%.

Nace also tested quantization to show the model travels well. It reports identical answers for bf16 and Q8_0 on an AWS g5.2xlarge (A10G, 24 GB), and says the Q8_0 GGUF matched on an Apple M5 Pro on both Metal and CPU. Truncating requests to 8K tokens, however, drops accuracy to 76.5% and 78%, so context length clearly matters to the result.

Integration is pitched squarely at coding agents. A Drex agent skill plugs the model into Claude Code, Codex, Cursor, OpenCode, Hermes Agent, Gemini CLI and GitHub Copilot, and Nace's launch post offers a $25 sign-up bonus for cloud users. Local deployment through Ollama and llama.cpp requires Nace's forks rather than mainline builds, and the weights ship under a RAIL-M license with use restrictions.

The open-source release is effectively a bet that decision-making deserves to be treated as a distinct layer in the agent stack, separate from the language model that does the talking. Nace has published the weights on Hugging Face alongside a GitHub repo and a product page, with the caveat that training on the benchmark's training splits may flatter familiar decision types and that results on new domains may differ.