A large share of “LLM tasks” in production write neither prose nor code: should this ticket be escalated, which of two summaries is more relevant, is this content compliant, does this candidate pass screening. The default answer of recent years is a long chain-of-thought generation followed by parsing — slow, expensive, and sometimes the answer flips when the options are merely reordered. The Drex family from Nace.AI takes a different route: pass in a context and a set of options, and a single forward pass returns a probability for every option, with no text generated at all. A 51-page technical report (arXiv:2611.07854) documents the methodology, ablations and deployment details, and all four sets of weights are open.
Decision models: a different readout, not a bigger model
The term follows the System One interface introduced by Jev: given a state (a document, record or agent context, optionally with images) and questions (multiple choice, yes/no, or a rating scale), the model returns a probability distribution over the options instead of natural language. The report argues that discriminative readouts patch three pain points of autoregressive pipelines in production: inconsistent answers, outputs that must be parsed, and cost and latency that grow with generation length. Classification, scoring and embedding fix their output shape at deployment time; decision models tackle the fourth kind, where the candidate set is supplied by the user at inference time — the shape most business decisions actually take.
One interface, four backbones
Drex is a family rather than a single model. DLM (8B) builds on NVIDIA Efficient-DLM-8B; v1.5 (9B) builds on MiMo-V2.6-Distill-Qwen-9B; v2.0 is a 35B mixture-of-experts with about 3B active parameters on the Qwen3.6-35B-A3B backbone, scoring 61.25 on the public Decision Index 0.3.1 when it shipped on October 13; the flagship Max is a dense 27B multimodal model trained from Qwen3.8-27B, released October 14. All four share one request format but split into two architectural camps: the pointer models (DLM, v1.5, v2.0) encode the context once and share it across questions, while Max re-encodes the full state for every question, removes the causal attention mask in its full-attention layers, and maps the final hidden state to logits over 255 option codes through a bias-free linear layer — costlier per question, in exchange for accuracy.
A three-stage recipe, with ablations to match
Max trains in three stages. Stage one is supervised fine-tuning with two consistency penalties: the predicted distribution must not change when the options are permuted, nor when the sentences or fields of the state are shuffled. Stage two is RLAF — reinforcement learning from adjusted feedback — where the reward blends the verified correct answer with the probabilities a judge model assigns to the options, and a KL term keeps the policy close to its own SFT checkpoint. Stage three tunes the image encoder and its projection so the model can decide directly from images. The ablations are concrete: RLAF lifts the Decision Index score by 2.0 points; the consistency terms cut the option-order flip rate from 6.3% to 4.2%; and the vision stage does not reduce accuracy on text-only inputs.
Scores and latency: strong on the report numbers, sober on the independent index
On the report numbers (Decision Index 0.4.1 public suite, leaderboard of October 22, 2026), Max scores 67.52 against 60.70 for Jev 1.14.0 — a new best among the decision models compared in the report — and holds 66.60 on maintainer-run private tests. The report also faces the independent view: on the Artificial Analysis Decision Index, Max ranks second among decision models with 64, versus 65 for the three leading models. On Decision Bench v4 it is statistically tied with d3-turbo and GPT-6 Sol and trails Claude Opus 5.5 by 1.6 points. Latency is the reason this model class exists: a typical request to Max completes in 183 ms, against roughly 2.0 s for GPT-6 Sol and 3.5 s for Claude Opus 5.5.
Openness and deployment
The code is Apache-2.0; Max, v2.0 and v1.5 weights ship under the Nace.AI Open RAIL-M license, while DLM follows the CC BY-NC 4.0 license of its backbone. Beyond the Python reference server there are llama.cpp and Ollama forks, a Homebrew-installable runner and MLX: the affine-Q4 checkpoint runs Drex Max on Apple Silicon Macs with 24 GB of memory, Q8_0 needs 48 GB, and Max supports contexts up to 262,144 tokens. Three models are also served through a hosted API and OpenRouter; at writing, the GitHub repository sits at 84 stars and 57 commits. The report even discloses its inference surgery: swapping hand-written Triton kernels for the FlashInfer Blackwell prefill kernel cuts latency by 9–16% on 8k–128k-token states, with 98.7% of top options agreeing with the previous kernels; quantizing all 128 MLP projections to NVFP4 shaves roughly 8% off median latency, but it landed 2.00 points above the BF16 reference on When2Call — past the 1.5-point-per-dataset limit the team set for itself. They shipped it after manual review and reported the difference.
Where to stay skeptical
The report admits its own limits: outside one controlled ablation on Max, most comparisons are associations rather than causal conclusions, and which design changes actually drive the gains remains open. Max re-encodes the full state for every question, which costs more than the pointer models on long multi-question contexts, and a 4.2% flip rate is not zero, so order-sensitive production use still needs headroom. Meanwhile Microsoft, Amazon, Cloudflare and Shanghai AI Lab have all pushed decision models in the past month, and Drex bundles a family, open weights and a full methodology into one release. While the industry piles compute onto generative reasoning, it is worth counting how much of the “generation” in your application is really just “selection”.