Drex Max: Nace's hosted multimodal decision model reads up to eight images per request
publicado (OK do Fabricio 20/10 BRT).
Nace.AI lists Drex Max (drex-max) in its models table with a release date of 13 October 2026; MarkTechPost covered it on 19 October. It is the most accurate Drex model by Nace’s account and the only one that reads images.
What the docs say. Drex Max takes the same request as the other Drex models on POST /v1/systemone. It is a pinned id: drex-latest still points to drex-v2.0, which also answers requests that omit model. Price is $0.06 per million input tokens, with output not billed. Limits: 16 questions per request, 255 options or score levels, 262,144 tokens per question (state plus question), up to 8 images per request and a 4 MiB body. It may shorten a long state before inference. It is hosted only; no weights or quantized builds are announced.
Vendor numbers (UnverifiedClaim). As reported by MarkTechPost from Nace’s own page: 67.59 on the public Decision Index 0.3.2, 92.2 on JevBench against 86.7 for Jev 1.14, 98.2 on BFCL tool calling and 0.780 on OSWorld-Verified; weakest are HLE (19.0) and MMLU-Pro (72.5). Nace says Drex trains on the index’s official training splits and is scored on held-out splits.
Independent boards (checked 20 Oct). Decision Index 0.4 (published 20 Oct; Full from the private parts only) lists “Nace.AI Drex Max v1 (27B)” at Full 68.51, #8, public 67.52, against 61.25 public for Drex v2.0. We did not find Drex Max on the JevBench API or open-weights boards.
Score. The Drex page is scored on the open drex-v2.0 weights and is unchanged; Drex Max is added there as a hosted variant.