Nace.AI opens Drex v1.5: a 9B long-context decision model under a revenue-capped license
Nace.AI published Drex v1.5 with code on GitHub, a Q8_0 GGUF and a /v1/systemone server. It reads text or JSON up to 131K tokens and answers Choice, Noul and Score questions with a pointer head on MiMo-V2.6-Distill-Qwen-9B.
- Board: official Decision Index 0.3 full score 50.84 (rank 25, tie group 28); public part 58.08, the figure Nace headlines as “#1 under 10B” (UnverifiedClaim for the ranking claim).
- License: Nace AI Open RAIL-M. Free for users under $1M gross revenue; larger companies may use it for personal or research purposes only and need a commercial license otherwise.
- Sibling: Drex DLM, a diffusion-backbone variant, is CC BY-NC 4.0.
Catalog page: Drex. Those scores were applied on 10 Oct. On 12 Oct the page moved to Drex 2.0 at 6.6 / 7.3 / 4.
Update — OpenRouter and benchmark context (10 Oct 2026)
OpenRouter now lists Drex v1.5 through DeepInfra at $0.04/M input, with free output. Nace’s direct API is $0.05/M input, also with free output. OpenRouter’s page for the Nace-linked Drex app reports 1.76B total tokens in the current snapshot. This is app-level usage reported by the platform, not a named external production case.
Nace’s model card reports 58.08 for its own run of the public Decision Index 0.3.1 kit against Jev’s listed 57.96. The board describes the pair as within its 0.9-point tie band; this is not a statistical significance test, and the public-only score is distinct from Drex’s official full 0.3 score of 50.84. The 89.5% / 93.4% long-context results are also publisher-run. Nace’s separate 0.2.1 page reports GPQA Diamond 45.4% vs Jev 78.6%, a task-specific result that illustrates the trade-off in broad knowledge benchmarks.