Nace open-sources Drex 1.5, a model that scores choices

Nace AI posted on October 9, 2026 that it was open-sourcing Drex 1.5, calling it “the smallest #1 decision-making model,” built for local and cloud deployment with a 128k context window. The weights are in the public Hugging Face repo nace-ai/drex-v1.5.

Drex does not write text. The model card says it reads a state — a string, object, or list — and typed questions, then returns a probability for every supplied option. Nothing is generated. Each question is one forward pass over the state plus that question, so temperature, top_p, and top_k do not apply. An agent that only needs a probability over a fixed menu of options can run that scorer locally instead of calling a token-generating model.

The card documents three question types. choice takes named options and returns the pick plus a probability per option. noul is yes/no and returns the probability of yes. score takes an ordered scale, lowest first, and returns a probability-weighted score. Requests go to POST /v1/systemone, a TypeSafe-compatible format the card says matches the hosted Drex API and the open Drex DLM. The local server needs no Authorization header and, the card says, should stay on 127.0.0.1 unless you add authentication.

The listed backbone is Qwen3_5ForCausalLM: 32 layers, hybrid attention (three linear-attention layers for every full-attention layer), about 9B parameters (8.95B). A pointer head, head.pt, scores options from the backbone’s hidden states. The named base model is XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. bf16 weights are about 18 GB; a Q8_0 GGUF is about 9.5 GB.

Those figures do not all describe the same context limit. The launch post markets 128k. The card accepts 16,384 tokens by default and up to 131,072, and reports long-document accuracy out to 128k tokens.

Local runners on the card are a CUDA Python server (inference.py and serve.py; torch and transformers), a drex-v1.5 branch of Nace’s llama.cpp fork, and a drex-v1.5 branch of Nace’s Ollama fork. Apple silicon and CPU are directed to the forks, not the Python path. The October 9 post also names Unsloth, OpenCode, Claude, Codex, Cursor, Hermes, and Claw. That compatibility list is the post’s claim; the card does not restate those clients as tested runtimes.

Treat the rank as Nace’s claim. On the card’s Decision Index 0.2.1 table, Drex v1.5 is first at 58.28, ahead of Jev 1.13.0 at 57.91, and the card says it leads Jev on 21 of 38 index benchmarks. On the card’s own JevBench table (231 public items), Jev is slightly higher: 87.0% versus 86.2%.

Weights use the Nace.AI Open RAIL-M license. The bundled Kev scoring code is Apache-2.0.

Read the Hugging Face model card for the request JSON and the Python, llama.cpp, and Ollama steps. Use the October 9 post for the client list Nace named at launch, and do not treat the index rank as an independent result.