Drex v1.5

Drex v1.5, released September 28, 2026 by Nace.AI, is a 'decision model' rather than a chat model: it reads a document plus typed questions and returns a probability for every option, without generating any text. The open weights total roughly 9B parameters in bf16 (~18GB on disk), built on a Qwen3.5-based backbone distilled from Xiaomi's MiMo-V2.6, with 32 layers and hybrid attention (three linear-attention layers per full-attention layer). A pointer head scores options from backbone hidden states. It accepts 16K tokens by default and up to 131K, serving a TypeSafe-compatible API.

The honest picture: this is a niche architecture for classification-style decisions over long documents — it does not generate text at all, which limits its generality. Nace.AI is a small vendor, so ecosystem support, documentation depth, and long-term maintenance are real risks, and the 'decision model' framing is marketing for what is essentially a scoring head.