Nace AI open-sources Drex decision models, hitting decision-task SOTA at 183ms latency
m4rkGS · reddit · 2026-11-07
Nace AI published a 50-page technical report on Drex, a family of four open decision models (8B–35B) that do discriminative inference—outputting option probability distributions without autoregressive decoding. Training combines full-parameter SFT on 1.15M verified decision records, option-order consistency training, RLAF (RL from adjusted feedback), and confidence calibration; ablations show most gains come from SFT. Drex Max sets a new SOTA on Decision Index 0.4.1 (67.52), answers typical requests in 183ms (vs 2.0s for GPT-6 Sol, 3.5s for Claude Opus 5.5), and stays stable under option reordering. Weights ship in safetensors and GGUF for llama.cpp/Ollama, targeting agent routing and tool selection.
Related event: Nace AI Open-Sources Drex Decision Models with Direct Probability Output(8 posts)→
More from coding & agent
- One folder that Claude reads first: the underrated Claude setup for solo operators — PrajwalTomar_ · 2026-11-07
- The AI agent testing landscape: 7 tools mapped across 5 distinct testing jobs — njyx · 2026-11-07
- AI rotoscoping workflow today: LTX 2.5 Alpha LoRA + SAM3 + BiRefNet, layered ProRes — PurzBeats · 2026-11-07
- Coding agent mistook brainstorming for orders and deleted the monthly spending cap — Dapper_Pass26 · 2026-11-07
- Gnomancer: an open-source CLI and MCP server to explore your raw DNA data with AI agents — Ok_Mirror_832 · 2026-11-07
- Workflow tip: compile your health records into CLAUDE.md and let Opus analyze them — AaronBergman18 · 2026-11-07