Nace AI releases four decision models in under four weeks

single source· 4 articles · confidence: low · first seen 2026-10-10 04:51 UTC

What this means for you

The 27B weights are the only part you can act on today: if you already self-host, it is worth a look on a Mac. The 67.52 is Nace's own figure, published without a benchmark name, so it is not a comparison to anything you run. Nothing here should change a routing decision.

Nace AI has shipped four Drex models in under four weeks — separate releases, not one launch — ending with an open-weights 27B version that runs on Apple Silicon. The company reports 67.52 for Drex Max v1, with no benchmark name, evaluation date or harness given. Earlier came the open-sourced 9B Drex 1.5 on 9 October, a multimodal Drex Max at 67.59 on 19 October, and Drex for Search, reported at 92.05% on BrowseComp-Plus using 12.6 search calls. No score here has been checked by an outside evaluator.

Key facts

  • ·Drex Max v1 is a 27B multimodal decision model released as open source on 1 November 2026. source
  • ·Nace reports Drex Max v1 scoring 67.52, with no evaluation date or harness given. source
  • ·Drex for Search is reported at 92.05% on BrowseComp-Plus using 12.6 search calls. source
  • ·Drex 1.5 is a 9B decision model released open-source on 9 October 2026. source
  • ·Drex Max, a multimodal decision model for computer-use agents, is reported at 67.59 on 19 October 2026. source

What the sources say

  • MarkTechPost — Introduces the 9B open-source model and the premise of ranking candidate options rather than writing answers.
  • MarkTechPost — Covers the multimodal variant for computer-use agents and its 67.59 reported score.
  • MarkTechPost — Details the search variant, with retrieval numbers on BrowseComp-Plus and turn counts on OfficeQA Pro.
  • MarkTechPost — Announces the 27B weights release and local inference on Apple Silicon.

Sources

The original reporting. Follow these — they did the work.

← the wire