Nace AI 发布多模态决策模型 Drex Max，Decision Index 得分 67.59
Meet Drex Max: Nace AI’s Multimodal Decision Model for Computer-Use Agents with a Score 67.59 - MarkTechPost
Nace AI 发布 Drex Max，这是其首个多模态决策模型，可读取文本和图像并返回校准后的概率，不生成文本。该模型在公开 Decision Index 0.3.2 上得分 67.59，比 TypeSafe 的 Jev 1.14 高 6.9 分；单次截图决策中位延迟 174 ms，输入定价为每 1M tokens 0.06 美元，输出不计费。模型仅通过 Nace Console API 提供，参数规模未披露、权重闭源，上述分数均为 Nace 自报。
正文
当前显示原文，中文全文译文尚未就绪。
Add as a preferredsource on Google
Nace AI has released Drex Max, a multimodal decision model for agents that use computers and browsers. The Drex Max multimodal decision model reads text and images and returns calibrated probabilities. It writes no text. Nace team reports 67.59 on the public Decision Index 0.3.2, 6.9 points above TypeSafe’s Jev 1.14. For agent builders, that means one cheap, typed call per click.
TL;DR
Size: Parameters not disclosed. Up to 262,144 tokens per question on the API, including image tokens.
Runs on: Hosted API only (drex-max on Nace Console). No open weights or quantized builds announced.
Performance: 67.59 on Decision Index 0.3.2 (public), the top score Nace lists for any decision model.
Best: 98.2 on BFCL tool calling, and 92.2 on JevBench versus 86.7 for Jev 1.14.
Worst: 19.0 on HLE, and 72.5 on MMLU-Pro versus 82.1 for Jev 1.14.
What is Drex Max?
Drex Max is a decision model from Nace AI that answers typed questions about text and images. It is the most accurate model in the Drex family, and the only one that reads images.
A decision model does not generate tokens. You send a state and named questions. Each question is one of 3 types: noul (yes/no), choice (pick a label) or score (ordered scale). The model returns a probability distribution for each answer in 1 HTTP call. Drex can only answer with the options you gave it.
How does Drex Max work with images?
Drex Max uses the same POST /v1/systemone endpoint as every Drex model. You select it with "model": "drex-max". It is a pinned ID, so drex-latest does not route to it.
Images go in an images array. Every question sees the same images, before the state and questions. You refer to them as ‘Image 1’ and ‘Image 2’ in plain text. Each entry is either an uploaded workspace file or inline base64.
Key image rules from the Drex Max guide:
Formats: still PNG, JPEG or WebP.
Up to 8 images per request, 16 MiB each, 32 MiB total.
Each image is resized to at most 1,638,400 pixels (1,280 × 1,280).
Each image costs up to 1,602 tokens in every question.
Up to 16 questions per request, with a 4 MiB request body.
How does Drex Max perform on benchmarks?
All scores below come from Nace’s Drex page. They are self-reported. Nace says Drex trains on the index’s official training splits and is scored on held-out splits.
Decision benchmarks: Drex Max scores 67.6 on the public Decision Index 0.3.2. Jev 1.14 scores 60.7 and Drex 2.0 scores 61.3. On JevBench, Drex Max reaches 92.2 against 86.7 for Jev 1.14.
Agentic benchmarks: Drex Max posts 0.780 on OSWorld-Verified and 98.2 on BFCL. It scores 85.2 on When2Call and 84.8 on RAGTruth.
Vision: On the Decision Index Vision public suite, Drex Max scores 74.5. GPT-6 Luna Decisions scores 63.9. Mind2Web reaches 88.6 and InfographicVQA reaches 98.8.
Where it trails: Jev 1.14 leads on MMLU-Pro (82.1 vs 72.5), HLE (19.4 vs 19.0) and ARC-Challenge (97.6 vs 97.4). On Hateful Memes, Drex Max trails GPT-6 Luna Decisions, 70.6 vs 71.2.
How fast and how cheap is Drex Max?
Nace lists a 174 ms median per screenshot decision. Jev 1.14 takes 253 ms on the same 243 questions. Nace says Drex Max accepts every screenshot size, while Jev 1.14 rejects 15.9% by size.
Across its benchmark table, Nace reports a 183 ms median latency. The launch post cites 34 ms per decision and under 1 second at 262K tokens. Those figures use different setups, so treat them as separate claims.
Pricing is $0.06 per 1M input tokens. Output tokens are not billed. New accounts get $25 in signup credit, and failed requests cost nothing.
How does Drex Max compare with other decision models?
FeatureDrex MaxJev 1.14Drex 2.0d3-flash
ParametersNot disclosedNot disclosed35B total, about 3B active (MoE)8.39B
Context262,144 tokens per question64K per request, 262K with long-context header16,384 default, up to 131,072Not disclosed
InputsText, imagesText (TypeSafe docs)TextText, images, video
LicenseProprietary (hosted)Proprietary (hosted)Nace.AI Open RAIL-MApache-2.0
Decision Index (public)67.6 (0.3.2)60.7 (0.3.2, per Nace)61.25 (0.3.1)57.8 (0.3.2)
Vision index (public)74.5Not disclosedNot applicable69.9
Price per 1M input tokens$0.06$0.042$0.06Self-host, free trial
Runs onNace Console APITypeSafe APISingle accelerator, bf16 about 70 GB1 AMD Instinct MI325X tested
Note: d3-flash’s vision figure is its own evaluation on the 0.3.2 vision board. Drex Max’s figure is Nace’s own run. Cross-vendor numbers are not a controlled comparison.
Key Takeaways
Drex Max is Nace AI’s first multimodal decision model.
It scores 67.59 on Decision Index 0.3.2, 6.9 above Jev 1.14.
Screenshot decisions take 174 ms, versus 253 ms for Jev 1.14.
Pricing is $0.06 per 1M input tokens, with output free.
Size is undisclosed, weights are closed, and scores are self-reported.
Michal Sutter+ postsBio ⮌
Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
Michal SutterExa 3.0 Search API Tops Perplexity, Brave and Parallel on 8 Benchmarks
Michal SutterMeet Grok Voice 3.0: SpaceXAI Voice Agent Model Ranks #1 on Artificial Analysis
RELATED ARTICLESMORE FROM AUTHOR
Exa 3.0 Search API Tops Perplexity, Brave and Parallel on 8 Benchmarks
MiniMax Just Launched M3.1 Flash: 428B MoE Agentic Model at $0.15/M Input
NVIDIA Open Sources Run:ai Under Apache 2.0 for Kubernetes GPU Orchestration