Nace AI Open-Sources 27B Drex Max Decision Model
The open-weight model returns decision probabilities in a single pass and can run on consumer Apple Silicon Macs and NVIDIA GPUs.
Updated on Nov. 1, 2026 in Artificial Intelligence
Live Poll
Do you think AI tools running on your own computer will be good for you?
Nace AI has released Drex Max v1, a dense 27-billion-parameter decision model, as open weights. The model ships in several formats, including bf16 weights, a Q8_0 GGUF, and an affine Q4 checkpoint, along with forks of llama.cpp and Ollama.
Why it matters
Releasing the weights lets developers run a model of this size locally on Apple Silicon Macs and NVIDIA GPUs rather than relying only on hosted services. The release also gives researchers a public benchmark and a model they can inspect and adapt.
Drex Max scores 67.52 on the public Decision Index 0.4, 6.27 points ahead of Drex v2.0, and reaches 97.3 on InfographicVQA and 81.9 in its Tools area. The source does not report results on benchmarks beyond these.
The players
Nace AI
Nace AI is the developer that released Drex Max v1 under its Open RAIL-M license.
The details
Drex Max is a dense model with a 262,144-token window that returns a probability for every option in one forward pass without generating text. Its architecture combines 48 Gated DeltaNet linear-attention layers with 16 full-attention layers placed every fourth layer.
Timeline
Nace AI released Drex Max v1 as open weights on November 1, 2026.
The Tech Race
Drex Max extends the Drex line, outscoring Drex v2.0 by 6.27 points on the Decision Index 0.4 while shipping as open weights rather than a closed service. The release positions Nace AI as an open-weight alternative in the decision-model segment.
Users with 24 GB or more of memory on an Apple Silicon Mac can run the affine Q4 checkpoint, while the Q8_0 GGUF requires 48 GB. On a 48 GB Apple M5 Pro, the Q8_0 GGUF matched H100 bf16 within 0.02 per probability, showing local hardware can approximate datacenter output.
The takeaway
Developers with capable local hardware can now test a 27B decision model without sending data to a hosted service. Checking the memory requirements for each checkpoint before downloading will avoid wasted time.
Further reading
Explore more on Artificial Intelligence to see how open-weight models are changing local deployment.
Source note: This article includes information reported by MarkTechPost.
Live Poll
Do you think AI tools running on your own computer will be good for you?
