Nace.AI open-sources Drex-Max v1: a 27B decision model with no token decoding
m4rkGS · reddit · 2026-11-01
Nace.AI released Drex-Max v1, an open-weight (Open RAIL-M) dense 27B multimodal "decision" model. Unlike LLMs, it generates no tokens: a 255-way readout head returns probabilities over candidate answers in a single forward pass.
- 64-layer hybrid backbone: 48 Gated DeltaNet linear-attention layers plus 16 bidirectional full-attention layers (one every 4th layer)
- Context up to 262K tokens (16K default); images up to 4.2M pixels, encoded once per batch
- Scores 67.52 on their structured decision benchmark, beating the larger 35B Drex v2.0 (61.25); Vision Index 73.92 (InfographicVQA 97.3, KIE 92.4)
- Runs 400ms per decision on an M5 Pro (Q80 GGUF, 27GB) within 0.02 of H100 bf16; CUDA bf16 needs 48.6GB VRAM, 8-bit 25.9GB
- Ships with a Python SDK and llama.cpp/Ollama forks for Apple Silicon, Metal, and CPU local inference
More from coding & agent
- Thorsten Ball: drop branches and PRs, push to main — the industry is catching up — RealGeneKim · 2026-11-02
- Jez Humble's talk: making CI/CD work for fully agentic delivery — RealGeneKim · 2026-11-02
- Super-agent context design cuts task time to 12.4 min vs 20.4 min baseline — Hesamation · 2026-11-02
- Jez Humble: everyone is rediscovering trunk-based dev and real CI — RealGeneKim · 2026-11-02
- The AI plan looked perfect — so it built a fish tank — jimmykoppel · 2026-11-02
- State-diagram skill for Claude Code: project memory turns gotchas into explicit flow conditions — DJIRNMAN · 2026-11-02