Nace AI opens Drex Max v1, a 27B decision model that runs on your own Mac

▶ Watch this story in one minuteThe same video we post on Instagram and TikTok

What you could do with it

• A company could run routing, classification, and policy-check calls locally on Apple Silicon Macs to keep sensitive data off the internet and avoid per-call cloud API fees ¹.
• A developer could integrate the model into applications via a common POST /v1/systemone API for fast decision tasks like triaging support tickets or enforcing guardrail rules ¹.
• An enterprise accounting team could automate back-office and front-office operations end-to-end using Nace's specialized models at one-tenth the cost of large language models ³.

What Drex Max actually does

Drex Max is a 27-billion-parameter dense model with a 262,144-token context window, though it defaults to 16,384 tokens per request ¹. Unlike a chatbot, it does not generate words. You send it a state — text, JSON, or images — together with named questions, and it returns a probability for every option in a single forward pass . Questions come in three flavours: multiple choice among named options, yes-or-no, and an ordinal score .

That design makes the model useful for tasks where an agent or application must decide where a request goes next, classify a support ticket, or enforce a guardrail rule. The company says the hosted version already powered routing, classification and policy checks for its customers ¹. Now developers can run the same logic on their own machines .

Architecture and local runtimes

The model uses a mixed architecture: 64 layers in total, with 48 Gated DeltaNet linear-attention layers that process tokens left to right and 16 full-attention layers placed every fourth step that are non-causal, meaning every token can see the entire sequence ¹. A native vision tower converts images so every question in a request can reference them . The output layer is a 255-way readout head that turns the final hidden state into option probabilities .

Nace published bf16 weights, a Q8_0 GGUF checkpoint, an affine Q4 checkpoint, and modified forks of both llama.cpp and Ollama so users can serve the model through a common POST /v1/systemone API ¹. Affine Q4 runs on Apple Silicon Macs with 24 GB of memory or more; Q8_0 GGUF requires 48 GB . NVIDIA GPU users can run the bf16 version on an H100 or the 8-bit version on an L40S with 48 GB of VRAM . CPU inference is available through the llama.cpp and Ollama forks .

Performance that beats previous versions

On the public Decision Index 0.4 benchmark, Drex Max v1 scored 67.52, leaving the earlier Drex v2.0 behind by 6.27 points (61.25) and Jev 1.13.0 by 9.56 points (57.96) ¹. All figures are self-reported by Nace AI . The model also reached 97.3 on InfographicVQA and 81.9 in the tools category of its area skill breakdown . On Artificial Analysis's own Decision Index the model ranked 4th out of 18 tested models with a score of 64 . In JevBench it achieved 92.2%, and on Decision Bench v4 it hit 94.30% .

In Nace's tests, the quantizations performed reliably. On a 48 GB Apple M5 Pro, the H100 bf16 version and the Q8_0 GGUF model achieved accuracy matching within 0.02 probability, while the affine Q4 variant stayed within 0.0329 of Q8_0 and produced no choice flips. Local Mac and CPU runs are text-only for now; image support requires the GPU Python runtime ¹.

Pricing and availability

The weights are open under the Nace.AI Open RAIL-M license and are available on Hugging Face ¹². At launch the hosted API price is $0.03 per 1,000 decisions, down from $0.06 for Drex v2.0 and $0.04 for Drex v1.5 . The hosted API for Jev costs $42 per 1 million tasks . For developers who want to self-host on Apple Silicon, the smallest checkpoint is the affine Q4 at 14.45 GB . The model is best suited for teams that already run workloads on Macs or prefer to keep decision logic on-premises .

Key facts

• Drex Max v1 is a 27B dense decision model with a 262,144-token window, open-sourced by Nace AI under RAIL-M
• It scores 67.52 on Decision Index 0.4, outperforming Drex v2.0 (61.25) and Jev 1.13.0 (57.96)
• Affine Q4 runs on Apple Silicon Macs with 24 GB+; Q8_0 GGUF requires 48 GB; NVIDIA GPUs support bf16 and 8-bit variants
• Q8_0 on an M5 Pro matches H100 bf16 within 0.02 probability; affine Q4 stays within 0.0329 with no choice flips
• Hosted launch price is $0.03 per 1K decisions, half the cost of Drex v2.0

Sources

1. Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon — marktechpost.com
2. nace-ai/drex-max-v1 · Hugging Face — aicampaign.live
3. Nace.AI: Small models for professional work