What actually shipped on November 1
Drex Max itself is not new. Nace launched it in October as a hosted model, and at the time it would not say how large it was. What shipped on November 1 is the model as files: a dense 27B backbone, its 255-way readout head, and enough loaders and forks to run the whole thing without calling Nace's API. For a category built on hosted endpoints — TypeSafe's Jev is still closed and hosted — that changes who controls the decision layer inside an agent.
The practical shape is in the file sizes. The bf16 weights take about 49 GB. The build a Mac owner will reach for first, the affine Q4 checkpoint, is a 14.45 GB folder needing a Mac with at least 24 GB of memory. The Q8_0 GGUF is 27.25 GB and needs 48 GB. Nace documents five runtime paths using the System One request and response format. That gives developers a common interface for comparing hosted and local inference, although deployment still requires runtime-specific installation, configuration and hardware validation.
Nace specifies at least 24GB of unified memory for the affine Q4 checkpoint on Apple silicon and 48GB for the Q8_0 GGUF. On a 48GB Mac, its Q8_0 instructions recommend a 2,048-token context. The model's stated maximum is 262,144 tokens, but the default per-request setting is 16,384 and practical limits depend on the runtime and available memory. These are Nace's documented configurations, not independent compatibility guarantees.
The documented Mac builds are text-only
Drex Max can process images through its native vision tower, but that capability is not included in every distributed format.
The Q8_0 GGUF and affine Q4 checkpoints omit image support. Nace documents llama.cpp, Ollama, Kapitsa and MLX as text-only paths, including the available Apple silicon and CPU configurations.
For image inputs, Nace documents its Python loader and reports testing on NVIDIA H100 and L40S hardware. That is a statement about the supported and tested implementation, not proof that image inference is technically impossible on other hardware.
A developer downloading the Q4 checkpoint for a Mac can run local decision scoring on text and structured inputs, but cannot use that checkpoint to score screenshots or other image inputs. The distinction is material for computer-use agents that depend on visual state.
The published scores predate the weights release
Nace's model card dates its performance results October 19, 2026. The November 1 weights release does not introduce a new benchmark run.
Nace reports 67.52 on the public Decision Index 0.4, compared with 61.25 for Drex v2.0 and 57.96 for Jev 1.13.0. It also reports a 73.92 vision index, 92.2% on JevBench and 94.30% on Decision Bench v4.
The card lists a score of 64 on the Artificial Analysis Decision Index, ranking Drex Max fourth among 18 models at the stated snapshot. That is a separately identified benchmark, although the ranking and comparison presented here come from Nace's model card.
Nace says its training includes official training splits from the Decision Index suite, with evaluation on held-out splits. That is relevant context when interpreting the public score. It does not, by itself, establish evaluation leakage or invalidate the result.
The card also reports that Q8_0 probabilities on a 48GB Apple M5 Pro differed by no more than 0.02 from a bf16 H100 run on one example request. That is a useful compatibility check, not evidence of equivalent accuracy across a representative workload.
Open weights, restricted commercial rights
Drex Max v1 is released under Nace.AI's modified Open RAIL-M license, not Apache 2.0 or MIT. Its own LICENSE file permits certain commercial uses, but sets limits based on an organization's size and competitive relationship with Nace.
Under Attachment A, organizations with more than $1 million in gross revenue over the preceding 12 months, or more than $1 million in cumulative equity or debt funding, need a separate commercial license for uses outside the stated personal-use and research exceptions. The license also restricts use by organizations offering products or services that compete with Nace.
Other conditions prohibit fully automated decisions with legal or similarly significant effects on individuals without meaningful human review, alongside restrictions on discriminatory uses and regulated advice without appropriate oversight.
These are terms in the Drex Max v1 LICENSE file itself, not an inference from earlier Drex releases. The weights are publicly downloadable, but that does not make their use unrestricted. Commercial teams should review the actual license and their intended deployment before relying on the model in production.
Sources
- [1] Nace AI, Drex Max v1 model card, Hugging Face (read first-hand; benchmarks dated “As of 19 Oct 2026”; question types choice, noul and score; Q8_0 on a 48 GB Mac uses a 2,048-token context)Read source
- [2] Nace.AI Open RAIL-M License (Modified), the Drex Max v1 LICENSE file, Hugging Face (read first-hand; Attachment A paragraph 2: $1M preceding-12-months gross-revenue test, $1M cumulative equity/debt funding test, competitor restriction; paragraph 3: meaningful human review for fully automated decisions)Read source
- [3] MarkTechPost, “Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon,” November 1, 2026 (read first-hand; sponsored by Nace AI, as disclosed in the article; cited for release timing)Read source
- [4] MarkTechPost, “Meet Drex Max: Nace AI's Multimodal Decision Model for Computer-Use Agents,” October 19, 2026 (hosted launch; training-split statement)Read source