Drex 1.5 Review: A 9B Decision Model That Scores Options, Not Text

Nace.AI open-sourced Drex 1.5 on October 8–9, 2026: 8.95B dense parameters, about 18GB of bf16 weights; a Q8_0 GGUF of about 9.5GB runs on Apple silicon and CPUs. Default context is 16,384 tokens, up to 131,072. Three question types: choice, noul (yes/no) and ordinal score. It serves POST /v1/systemone, the same request format as TypeSafe's Jev; Nace says existing Jev clients migrate by changing a few environment variables. The backbone is Qwen3_5ForCausalLM, based on XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. License is Nace.AI Open RAIL-M: organizations with over $1M in yearly revenue or funding raised need a separate license beyond personal or research use.

The basis is model-card figures plus independent testing. Model card (Nace's own run of the official kit on the full public suite): 58.08 on Decision Index 0.3.1 public (Jev 1.13.0 at 57.96, both inside the board's 0.9-point tie band; ahead of Jev on 20 of 37 benchmarks); Full score 50.84 (25th of 117 rows, Jev at 60.11); 86.2% on JevBench's 231 items (Jev at 87.0%). Caveat: the public score is vendor-submitted and trained on the benchmarks' training splits, so the Full score (including private tests Nace could not train on) is the fairer number. Independent test (dev.to, October 10, 2026, 82 decision questions run locally on a Mac via OpenRouter): Drex 72/82, Microsoft Decision-1 75/82, Jev 79/82, Perplexity Decider 27B 80/82; Drex median latency 0.679s, $0.0030 per 1,000 calls.

1. Runnability: A single 24GB A10G runs bf16; the Q8_0 GGUF is about 9.5GB.
2. Long docs: 93.4% accuracy at 32k–128k context (model card), 2.0s median latency.
3. Weak spots: 45.4% on GPQA Diamond (Jev 78.6%); 7.4% per-review F1 on ACOS aspect sentiment.
4. Hosted: On OpenRouter via DeepInfra at $0.04 per 1M tokens.