Nace.AI releases Drex 2.0, a 35B-A3B decision model with open weights

Last updated:

Source · Model card: Drex

Nace.AI published Drex v2.0 on 12 Oct 2026: a mixture-of-experts checkpoint on Qwen3.6-35B-A3B (about 3B parameters active) plus a pointer head. The vision tower is dropped on load. It answers Choice, Noul and Score in one forward pass through POST /v1/systemone. Weights are bf16 (~70 GB). llama.cpp and Ollama need Nace’s forks on the drex-v2.0 branch.

  • Board: 61.25 is Nace’s own run of the public Decision Index 0.3.1 kit, against Jev 57.96 and Drex v1.5 58.08 on that card. It is not an official full-board line. Drex v1.5’s official full 0.3 score stays 50.84. Sieve 27B’s official 61.25 is a different result that happens to use the same digits.
  • JevBench, on the same card: Drex v2.0 85.7% (198/231), Jev 86.6% (200/231). Publisher-run.
  • Price: $0.06/M input on the Nace API (drex-v2.0) and on OpenRouter via DeepInfra. Output is not billed. v1.5 remains listed at $0.04/M on OpenRouter and $0.05/M on the Nace API.
  • License: Nace AI Open RAIL-M. Free, including commercial use, under $1M gross revenue; above that, personal and research use only.

The launch post also says inference is up to 38% faster than Drex 1.5 on 32K–128K inputs, and that Fireworks ships a dedicated deployment. Those two claims are not on the model card. OpenRouter’s provider is DeepInfra.

Catalog page: Drex. The quadrant point moves to v2.0 at 6.6 / 7.3 / 4.