Nace AI Releases Drex Max Decision Model as MiniMax Launches M3.1 Flash
Nace AI has released Drex Max, a multimodal decision model for computer-use and browser agents that reports 67.59 on the public Decision Index 0.3.2; MiniMax also launched M3.1 Flash with faster inference and lower agent-task costs.
Drex Max accepts a state and named questions, each defined as a yes/no "noul", a "choice" label, or a "score" on an ordered scale. It returns a probability distribution for each answer in one HTTP call. Nace says the model's parameters are not disclosed, and Drex Max is available only through a hosted API as "drex-max" on Nace Console, with no open weights or quantized builds announced. The API supports up to 262,144 tokens per question, including image tokens, and uses the same POST /v1/systemone endpoint as other Drex models. The model ID is pinned, so "drex-latest" does not route to it.
Image inputs go in an images array, and every question sees the same images before the state and questions. Users refer to them as "Image 1" and "Image 2" in plain text. Each entry can be an uploaded workspace file or inline base64. Supported formats are PNG, JPEG and WebP, with up to 8 images per request, 16 MiB each and 32 MiB total. Each image is resized to at most 1,638,400 pixels, or 1,280 by 1,280, and costs up to 1,602 tokens in every question. A request can contain up to 16 questions and a 4 MiB request body.
Nace's Drex page lists the benchmark scores, which are self-reported. Nace says Drex trains on the index's official training splits and is scored on held-out splits. The page lists 67.6 on the public Decision Index 0.3.2 for Drex Max, versus 60.7 for Jev 1.14 and 61.3 for Drex 2.0. On JevBench, Drex Max reaches 92.2 against 86.7 for Jev 1.14. On agentic benchmarks, it posts 0.780 on OSWorld-Verified, 98.2 on BFCL, 85.2 on When2Call and 84.8 on RAGTruth. On vision tests, it scores 74.5 on the Decision Index Vision public suite, compared with 63.9 for GPT-6 Luna Decisions, and reaches 88.6 on Mind2Web and 98.8 on InfographicVQA.
Drex Max trails Jev 1.14 on MMLU-Pro, 72.5 versus 82.1, on HLE, 19.0 versus 19.4, and on ARC-Challenge, 97.4 versus 97.6. On Hateful Memes, it trails GPT-6 Luna Decisions, 70.6 versus 71.2.
Separately, MiniMax launched the M3.1 Flash model on Oct. 19, according to Shaoshupai. The report says online inference speed is nearly three times faster than before, and comprehensive cost in Agent tasks is reduced by 70%. Space Bunny, an anonymous model that processed more than 200 trillion tokens, was an early version of M3.1 Flash. The model is now available on MiniMax API, MPlan and MiniMax Code, and its weights are planned for release later.
MiniMax says M3.1 Flash performed well on AutomationBench-AA, scored 76.2 on BankerToolBench and 73.7% on the software development benchmark DeepSWE v1.1. On computer use, it scored 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam, both first among domestic Chinese models according to the company. It also scored 82% on CyberGym, covering code auditing, vulnerability discovery and verification. The model targets complex agent tasks, software engineering and cross-application computer operation. In enterprise workflows, MiniMax says it can call tools across SaaS systems for sales, marketing, operations, customer service, finance and human resources; in software development, it can be used for code understanding, feature development, bug fixing and test validation; in computer use, it can interact with a real desktop environment. API pricing is 1 yuan per million input tokens, 4 yuan per million output tokens and 0.1 yuan per million cached tokens. MiniMax says a Pro version with a higher intelligence ceiling will follow.