AI BriefingNace.AI releases Drex Max v1, a 27B open-weight multimodal decision model with 262K context
·2026.11.02 00:31
Key point
The model uses a 255-way readout head to output probabilities over candidate answers in a single forward pass, avoiding token generation entirely.
Details
- Nace.AI open-sourced Drex Max v1, a dense 27B multimodal decision model under the Open RAIL-M license, designed for low-latency routing and agentic tool selection.
- Unlike standard LLMs, it does not generate tokens or JSON; instead, a 255-way readout head returns probabilities over candidate answers in a single forward pass with ~400ms latency.
- Benchmarks show Drex Max v1 (27B) outperforms the previous Drex v2.0 (35B) with a Public Decision Index of 67.52 versus 61.25, and achieves 94.30% on Decision Bench v4.
- Supports up to 262,144 tokens of context and 4,194,304 pixels per image, running on NVIDIA GPUs, Apple Silicon (Metal), and CPU via llama.cpp and Ollama forks.
- Hardware requirements include ~48.6 GiB VRAM for CUDA bf16 or ~27.25 GB for Q8_0 GGUF on Apple Silicon, with probabilities within 0.02 of unquantized bf16 on H100.
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
spinf-serve: Open-source inference server for bulk LLM scoring on single GPU