Nace AI opens Drex Max v1 weights: a 27B decision model that runs on a 24GB Mac, local paths text-only for now
According to a MarkTechPost report dated November 1, 2026 (noted at the end as sponsored by Nace AI), Nace AI released the weights of Drex Max v1, a 27B dense decision model, under Nace.AI Open RAIL-M, providing bf16, Q8_0 GGUF and affine Q4 checkpoints; it self-reports a Decision Index 0.4 score of 67.52.
What changed
According to MarkTechPost's November 1, 2026 report, which notes at the end that it was sponsored by Nace AI, Nace AI has released Drex Max v1 as open weights under the Nace.AI Open RAIL-M license. The report describes it as a 27B dense-parameter decision model: it does not generate text but takes a state (text, JSON or image) and a named question, returning a probability for each option in a single forward pass; questions fall into three types—choose one among named options, yes/no, and ordinal scoring—with up to 255 options per question. The model card describes a 64-layer backbone, of which 48 layers are left-to-right Gated DeltaNet linear attention and the other 16 are non-causal full attention placed every fourth layer, so each token can attend to the entire sequence; the config file lists a qwen3_5 model type and a 262,144-position window, with a default of 16,384. Release artifacts include bf16 weights, a Q8_0 GGUF (27.25 GB), an affine Q4 checkpoint (14.45 GB), and forks of llama.cpp and Ollama; the report says five runtimes uniformly expose a POST /v1/systemone interface. Benchmarks and quantization comparisons are all self-reported by Nace AI: Decision Index 0.4 scores 67.52, above the listed Drex v2.0 (61.25) and Jev 1.13.0 (57.96), with an Artificial Analysis Decision Index of 64; on a 48GB M5 Pro, Q8_0 and H100 bf16 differ by no more than 0.02 per option probability, and affine Q4 versus Q8_0 differs by no more than 0.0329 with no selection flips. The report also mentions a hosted version on OpenRouter, priced at $0.03 per thousand decisions at launch; current pricing should be checked against the vendor's pricing page.
What it means for developers
The first change for developers is the interface shape: this is not an interface you can substitute for a chat model just by turning prompts into options. Callers must first supply an option set and question type, and get back a set of probabilities; thresholds, abstention and fallback policies must be defined by the application, which is entirely different from parsing free-text output. The second change is deployment tiers: the memory thresholds in the report are affine Q4 for Macs with 24GB or more, Q8_0 GGUF needing 48GB, and on NVIDIA about 49GB for bf16 or about 26GB for 8-bit, meaning the same model can run on laptops, single GPUs and GPU servers, but with different precision and throughput. The third is a capability gap in local paths: llama.cpp, Ollama, Kapitsa and MLX do not support images; only Python with an NVIDIA GPU does, so ticket triage involving screenshots cannot be moved entirely to a Mac. The fourth is the nature of the license: downloadable weights do not equal an OSI-approved open-source license; Open RAIL-M carries use restrictions, and the terms should be checked before commercial use. The fifth is the gap in long-context claims: the window is listed as 262,144 but defaults to 16,384, and the report suggests a 2048 context for Kapitsa on a 48GB Mac, so the advertised long window is not the configuration you will actually run. It should be stressed that all the performance and quantization deviation figures above come from the vendor's self-report, and the source article is sponsored content, lacking third-party reproduction.
Integration and evaluation
The following are recommendations and conditional analysis based on the report, not verified facts. First, if your routing or triage workload is text-only, you could first try affine Q4 on a Mac with 24GB or more of unified memory: the report says the checkpoint is 14.45 GB with no selection flips, which provides a premise for offline canary testing, but only if the workload contains no images and you can accept single-machine inference latency; the evaluation criterion is to compare top-1 agreement and flip rate near the threshold between Q4 and bf16 or the hosted version on your own labeled historical tickets, and only switch traffic if agreement reaches a preset threshold. Second, if inputs contain screenshots or scans, the report says images go through a native vision tower and are visible to all questions in the request, but none of the four local paths supports images, so the choices are to keep GPU or hosted-side calls, or convert images to text first and then go local; the evaluation criterion is to test the vision path and an OCR-plus-text path on your own screenshot set for selection accuracy and unanswered rate, rather than looking only at the public vision index of 73.92. Third, for long inputs, set the context length explicitly and re-test at the target length: because the report says unanswered requests are counted as errors in the result dataset, unanswered cases from long context will directly raise the error rate, so evaluation metrics should include the unanswered proportion and accuracy decay curve at different context lengths. Fourth, the design of 255 readout heads answering multiple questions in one forward pass makes it possible to merge multiple policy checks into one request, provided the option sets of the merged questions do not conflict and the total length remains within the window; the evaluation criterion is single-request latency, per-question probability differences, and unanswered rate under long context for merged versus split approaches.
Limits and open questions
Drex Max v1 is not currently in KimiSeek's model catalog; the hosted API and cloud console availability mentioned in the report refer to Nace AI, and whether KimiSeek will offer it and at what price should be checked against the catalog and pricing page. All benchmarks, rankings and quantization deviation figures in the report are self-reported by Nace AI, and the source article is vendor-sponsored with no independent reproduction seen. The quantization comparison was done only on one 48GB M5 Pro, and figures such as a 0.0329 probability deviation cannot be directly extrapolated to other hardware and quantization combinations. The license is Nace.AI Open RAIL-M; open weights do not mean an OSI open-source license, and use restrictions should be checked independently. The qwen3_5 model type in the config only reflects a structural field; the report does not explain the training data source or base-model relationship. In addition, the report does not describe the local interface's specific behavior when no suitable option exists, only that unanswered cases are counted as errors in the result dataset.