Nace AI Open-Sources Drex 1.5 Decision Model

Nace AI has released open weights for Drex 1.5, an 8.95-billion-parameter dense decision model engineered specifically to evaluate pre-defined options instead of generating freeform text. Built on a distilled Qwen 3.5 9B backbone (MiMo-V2.6-Distill-Qwen-9B) featuring a specialized pointer head, the system processes state context alongside typed questions—such as choices, yes/no queries, or ordinal ratings—and outputs precise probabilities in a single forward pass.

In benchmark evaluations, Drex 1.5 achieved a public score of 58.08 on Decision Index 0.3.1, placing it within a narrow tie band alongside TypeSafe's closed Jev 1.13.0 model (57.96) and ahead of Bespoke Nimble 9B v3 (57.19). It performed particularly well on long-context processing, recording 93.4% accuracy on documents spanning 32K to 128K tokens with a median latency of 2.0 seconds, compared to 89.5% accuracy on 8K to 32K token texts. On JevBench, it reached 86.2% accuracy overall and matched Jev's 73.9% score on difficult items. In direct competition across eight OpenSpiel games, Drex recorded 122 wins, 47 draws, and 87 losses against Jev.

However, the model exhibits notable limitations in broad knowledge and fine-grained sentiment analysis. Drex 1.5 scored 45.4% on GPQA Diamond compared to Jev's 78.6%, 58.7% on MMLU-Pro compared to 82.7%, and registered a 7.4% F1 score on ACOS aspect sentiment versus Jev's 29.5%. Its capability distribution heavily favors tool selection, scoring 75.0 in Tools against 44.6 in Knowledge and Reasoning.

For practitioners, Drex 1.5 presents a cost-effective alternative for high-throughput decision systems. The model operates on a single 24 GB A10G GPU in bf16 precision or as a 9.5 GB Q8_0 GGUF file on Apple silicon and CPUs via specialized llama.cpp and Ollama forks. It serves the POST /v1/systemone API format, making it directly drop-in compatible with existing Jev integrations. Cloud deployments hosted via OpenRouter are priced at $0.04 per 1M input tokens with zero output costs under a Nace.AI Open RAIL-M license.