AI BriefingNace AI Research Lab publishes 50-page technical report on training SOTA decision models
·2026.11.06 23:39
Key point
The 50-page report details architecture, training, and evaluation for Drex models, which achieve 93.4% accuracy on long-context benchmarks.
Details
- Nace AI Research Lab published a 50-page technical report defining the architecture, training, and evaluation of their Drex decision models.
- Drex v1.5 achieves 93.4% accuracy on a 32K–128K token long-context benchmark, while Drex v2.0 maintains a 1.37% calibration ECE on approximately 115K held-out decisions.
- Training relies on full-parameter SFT with ~1.15M verified decision records, with ablation studies showing SFT drove most improvements while Reinforcement Learning from Adjusted Feedback (RLAF) provided smaller gains and better calibration.
- The model family ranges from 9B to 35B parameters, utilizing architectures such as Dense Causal Transformers, Diffusion LMs, and Mixture of Experts, with inference latency under 100 ms for 128K-token contexts.
- Weights for Drex DLM, v1.5, v2.0, and Drex Max are open-sourced in safetensors and GGUF formats, supporting llama.cpp and Ollama for local deployment.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Thinking Machines AI introduces Inkling 2 Small, planning Apache 2.0 open weights release next weekLongCat-2.5 Launches with Native Multimodality and 1.6T Parameters for Agentic Coding