Nace AI has published a 50-page technical report detailing the training methodology for its Drex decision model. Version 1.5 reached 93.4% accuracy in its long-context test, but the bigger development is that the full pipeline can finally be reproduced.
What this is
Drex is a family of open-weight models designed for decision-making tasks, with parameter counts ranging from 9 billion to 35 billion. These “decision models” are built to determine the next option, tool call or action path, rather than merely continue generating text as general-purpose models do.
Training takes place in three stages. First, the model undergoes supervised fine-tuning (SFT), using roughly 1.15 million verified decision records to learn from correct answers. Next, Nace adjusts option order and state arrangements. Finally, it applies “reinforcement learning from adjusted feedback,” or RLAF, combining outcomes with reviewer feedback for further optimization. The report also discloses the model architecture, evaluation sets, quantization methods and inference configurations. The weights can run locally through Ollama and llama.cpp.
Industry view
We believe this report is more valuable than another new leaderboard position. The open-source community has often seen model weights without the data construction, post-training recipe or evaluation boundaries behind them. As a result, enterprises have struggled to determine whether high scores come from the method itself or from a particular test setup. Drex at least puts these questions on the table.
Claims of the “first complete methodology,” however, still require caution. Nace conducted its evaluations internally and has yet to publish controlled comparisons of different architectures on the same underlying model. Problems such as weak closed-book knowledge and training costs across different data scales also remain unresolved. Its claim of maintaining a leading position “for several consecutive weeks” likewise rests mainly on internal testing and cannot automatically be equated with industry leadership validated by third parties.
Another notable conclusion is that most of the gains came from supervised fine-tuning. RLAF delivered comparatively limited improvements; its primary contribution was better confidence calibration, meaning the model's confidence in its answers more closely reflected its actual accuracy. This is a reminder that sophisticated training methods are not inherently superior to solid data and foundational alignment.
Impact on regular people
For enterprise IT: Decision models and tool selection are emerging as a distinct layer. Businesses can use Drex to assess whether they need private deployment, local routing or an in-house training pipeline, but should first validate third-party performance and results on real business data.
For workplace productivity: Open weights and local execution make it easier for technical teams to experiment with automated workflows. Nontechnical readers do not need to follow every training term; the core question is whether a model can reliably complete a specific task, not how many parameters it has.
For the consumer market: More reliable action selection could improve software assistants and local applications. Until the evaluations are independently reproduced, however, a figure such as “93.4% accuracy” should not be treated as a direct purchase criterion.