Nace.AI's Drex Max Scores Typed Questions With Probabilities From a Local 27B, Skipping Token Generation Entirely

Nace.AI has released Drex Max, a dense 27-billion-parameter model that answers typed questions about a given context by returning a probability for every option in one forward pass—and generates no text at all. The Hugging Face card for nace-ai/drex-max-v1 presents it as a multimodal decision model: pass context as state with one or more named questions, and receive structured scores instead of a free-form reply. Weights are published under the Nace.AI Open RAIL-M license, which is more restrictive than Apache-2.0 and should be treated as such by anyone planning redistribution or commercial deployment.

The model page’s performance snapshot is dated October 19, 2026. Code and documentation live in the nace-ai/drex-decision-models repository, with companion Q8_0 GGUF and affine Q4 checkpoints and forks of llama.cpp and Ollama on a drex-max-v1 branch.

A 255-way readout instead of a chat head

Drex Max is built around a 64-layer backbone: 48 Gated DeltaNet linear-attention layers that stay causal left-to-right, and 16 full-attention layers—every fourth layer—that are non-causal, so every token can attend to the whole sequence. Images go through a native vision tower and remain visible to every question on the request. A separate 255-way readout in readout.safetensors turns the last hidden state into probabilities over up to 255 options.

Question types are choice, noul (a graded yes/no-style output), and score. The bundled loader applies the non-causal attention mask; the card warns that a plain causal chat run will not reproduce the published scores. Default context is 16,384 tokens per request, raisable to the model’s 262,144-token window via environment flags or --max-length.

Leaderboard numbers Nace.AI reports

On the public Decision Index 0.4 text suite, Drex Max is listed at 67.52, ahead of Drex v2.0 at 61.25 and Jev 1.13.0 at 57.96. The Tools area column is 81.9. On the Artificial Analysis Decision Index, a weighted blend of AA-Omniscience-Decision, HLE-MultiChoice, When2Call, and WildGuardTest, Drex Max scores 64 and ranks fourth of 18 in the table Nace.AI reprints, one point behind three models tied at 65. The public vision index is 73.92, with InfographicVQA at 97.3 skill on 244 rows. JevBench is reported at 92.2% (213/231) and Decision Bench v4 at 94.30% (1010/1071).

All of these figures are company-published or drawn from boards the company cites. Independent replication is not claimed on the model card.

Where it actually runs

BF16 weights take about 49 GB; loading the backbone in 8-bit on CUDA drops that to about 26 GB. Nace.AI tested an NVIDIA H100 80 GB with bf16 and an L40S 48 GB with 8-bit; 8-bit probabilities on the example request stayed within 0.03 of bf16. On Apple silicon, the Q8_0 GGUF needs 48 GB of unified memory, while the affine Q4 checkpoint is aimed at Macs with 24 GB or more under Kapitsa and MLX. On an Apple M5 Pro with 48 GB, the company says Q8_0 on Metal answered its example request within 0.02 of the H100 bf16 run.

Python (python -m drex_max) is the path that supports images. Kapitsa, MLX, llama.cpp, and Ollama serve the text backbone through POST /v1/systemone with the same request and response shape, minus images. Building those runners means using Nace’s forks rather than stock upstream binaries.

Open RAIL-M, System One, and what the product is for

The license flag matters. Open RAIL-M carries use restrictions that ordinary Apache or MIT cards do not; operators should read the LICENSE file in the model tree before shipping a decision service on top of the weights. The API surface is deliberately narrow: one endpoint, typed questions, probability maps back. That makes Drex Max a fit for routing tickets, scoring urgency, gating tool calls, or classifying multimodal context where a generative model’s prose would be discarded anyway.

Nace.AI positions the release as a local alternative to hosted decision APIs that speak the same System One contract. For teams that already budget for a mid-sized GPU or a high-memory Mac, the practical question is whether probability-native scoring, with the Open RAIL-M caveats, beats wrapping a general chat model in a brittle JSON schema. The card’s answer is architectural: nothing is generated, every option gets a calibrated mass, and the vision tower is optional rather than bolted on after the fact.