Nace AI has launched Drex Max v1 as open weights. Drex Max is its most succesful resolution mannequin, and it now runs by yourself {hardware}. The discharge ships bf16 weights, a Q8_0 GGUF, an affine Q4 checkpoint and forks of llama.cpp and Ollama. For builders, Drex Max open weights imply routing, classification and coverage checks can keep on an area machine.
TL;DR
- Dimension: Dense 27B parameters, 262,144-token window (16,384 by default), as much as 255 choices per query.
- Runs on: NVIDIA GPU with Python (H100 bf16, or L40S 48 GB in 8-bit). Apple Silicon Macs with 24 GB+ (affine This fall) or 48 GB (Q8_0 GGUF). CPU by way of llama.cpp and Ollama.
- Efficiency: 67.52 on the general public Resolution Index 0.4, 6.27 factors forward of Drex v2.0 (self-reported).
- Finest: 97.3 ability on InfographicVQA. Instruments space reaches 81.9. A top-tier resolution mannequin now you can run offline on a Mac.
What’s Drex Max?
Drex Max is a call mannequin from Nace.AI that solutions typed questions on a given context. You move a state (textual content, JSON or pictures) plus named questions. The mannequin returns a likelihood for each choice in a single ahead move. Nothing is generated.
Questions are available in 3 varieties: alternative amongst named choices, sure/no, and an ordinal rating. That makes it a match for agent routing, ticket triage and guardrail checks. The hosted version is introduced at routing, classification and coverage checks.
How does Drex Max work?
The model card describes a dense 27B spine with 64 layers. 48 are Gated DeltaNet linear-attention layers that run left to proper. The opposite 16 are full-attention layers, positioned each fourth layer, and they’re non-causal. Each token can attend to the entire sequence.
Pictures move by a local imaginative and prescient tower and are seen to each query in a request. A 255-way readout head then turns the final hidden state into choice chances. The config lists a qwen3_5 mannequin sort with a 262,144-position window.
(operate(){var f=doc.getElementById(‘nx-drexmax-frame’);window.addEventListener(‘message’,operate(e){if(e.knowledge&&e.knowledge.sort===’nx-drexmax-height’&&f&&e.supply===f.contentWindow){f.fashion.peak=e.knowledge.peak+’px’;}});})();
How does Drex Max carry out on benchmarks?
All figures beneath are self-reported by Nace AI,
- Resolution Index 0.4 (public textual content): 67.52. Drex v2.0 scores 61.25 and Jev 1.13.0 scores 57.96.
- Space ability: Instruments 81.9, Language 72.2, Retrieval 68.8, Information and Reasoning 57.1, Arts 53.4.
- Public imaginative and prescient index: 73.92 throughout 11 benchmarks and 12,210 rows.
- Synthetic Evaluation Resolution Index: 64, which Nace says ranks 4th of 18.
- JevBench: 92.2% (213/231). Resolution Bench v4: 94.30% (1010/1071).
The results dataset covers 140,620 requests. Unanswered requests rely as unsuitable.
How do you run Drex Max regionally?
Nace paperwork 5 runtimes. All of them serve the identical POST /v1/systemone API.
| Runtime | {Hardware} | Checkpoint | Pictures |
| Python | NVIDIA GPU | bf16 (~49 GB) or 8-bit (~26 GB) | Sure |
| llama.cpp fork | Apple Metallic, CPU | Q8_0 GGUF (27.25 GB) | No |
| Ollama fork | Apple Metallic, CPU | Q8_0 GGUF | No |
| Kapitsa v0.2.25+ | Apple Silicon | Q8_0 GGUF or affine This fall | No |
| MLX | Apple Silicon | affine This fall (14.45 GB) | No |
The affine Q4 checkpoint targets Macs with 24 GB or extra. The Q8_0 GGUF wants 48 GB. On a 48 GB Mac, Nace recommends –context 2048 for Kapitsa.
Ollama customers set num_gpu 40 to dump 40 layers to Metallic. Setting num_gpu 0 runs it on the CPU. The Ollama fork wants a CAPABILITY resolution line within the Modelfile.
Quantization holds up in Nace’s checks. On a 48 GB Apple M5 Professional, the Q8_0 GGUF matched H100 bf16 inside 0.02 per likelihood. The affine This fall stayed inside 0.0329 of Q8_0, with no alternative flips.
Drex Max vs different resolution fashions
| Drex Max v1 | Drex v2.0 | Drex v1.5 | Jev 1.13.0 | |
| Parameters | 27B dense | 35B MoE, ~3B active | 9B | Not published |
| Max context | 262,144 | 131,072 | 131,072 | Not disclosed |
| Inputs | Text, JSON, images | Text (vision tower dropped) | Textual content | Not disclosed |
| Weights | Open, Nace.AI Open RAIL-M | Open, Nace.AI Open RAIL-M | Open, Nace.AI Open RAIL-M | Closed, hosted |
| Smallest native file | Affine Q4, 24 GB+ Mac | Q8_0, 36.9 GB | Q8_0, ~9.5 GB | Not relevant |
| Public Resolution Index | 67.52 (v0.4) | 61.25 (v0.3.1) | 58.08 (v0.3.1) | 57.96 |
| JevBench | 92.2% | 85.7% | Not disclosed | 86.6% |
| AA Resolution Index | 64 | 58 | Not listed | 58 |
| Hosted value | $0.03 / 1K decisions (launch) | $0.06 / 1K decisions | $0.04 / 1K decisions | $42 / 1M tasks |
Nace AI staff states that Resolution Index 0.4 makes use of the identical public suite as 0.3.1.
Key Takeaways
- Drex Max v1 weights at the moment are public beneath Nace.AI Open RAIL-M.
- It scores 67.52 on Resolution Index 0.4, above Drex v2.0 (61.25).
- A 4-bit construct runs on 24 GB Apple Silicon Macs by way of Kapitsa or MLX.
- Q8_0 on an M5 Professional matched H100 bf16 inside 0.02.
- Native Mac and CPU paths are textual content just for now.
Take a look at the model weights, Q8_0 GGUF, code and docs, Tweet and cloud console.
Because of the Nace AI staff for the thought management / assets for this text. This text is sponsored by Nace AI.
The put up Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon appeared first on MarkTechPost.

