MiniMax Launches New AI Model

MiniMax Group Inc., a Chinese artificial intelligence company based in Shanghai, has launched "MiniMax M3.1 Flash" on its API. The model is a type of artificial intelligence (AI) model called "mixture-of-experts" (MoE) and is offered at a price of 0.15 dollars per million input tokens.

Model's Performance Benchmarks

According to figures provided by the company, the model achieves a score of 62.2 on Automation-Bench v1.0.6, surpassing Opus 5 (50.3) and GPT-5.6 Sol (45.8). The model includes 428 billion parameters, of which about 23 billion are active, and it accepts input data in the form of text, images, and videos, returning text output.

Specialized Capabilities

The model also features a sparse attention mechanism and a native visual encoder. The context window is 1,000,000 tokens. In terms of coding, M3.1 Flash achieves a score of 73.7 on DeepSWE v1.1, slightly more than GPT-5.6 Sol (73.0), against 74.0 for Opus.

Cybersecurity and Vision Performance

In cybersecurity, the model achieves a score of 82.2% on CyberGym, demonstrating cutting-edge capabilities in detecting and validating vulnerabilities. In vision, MiniMax M3.1 Flash reaches 91.0 on BabyVision with tools and 82.5 on CharXiv RQ, but achieves a score of 43.4 on HLE without tools.

Model Availability and Use Cases

The model weights are expected to be available by the end of October 2026. MiniMax M3.1 Flash is intended for programming, automated administrative tasks, and visual reasoning. It is part of the six "Chinese AI tigers." The company was listed on the Hong Kong Stock Exchange in January 2026.

Benchmark Comparisons

On the CoWork benchmark, M3.1 Flash leads the Automation-Bench v1.0.6 ranking with a score of 62.2. It is tied with Opus 5 on BankerToolBench, with a score of 76.2. It also leads the LOCA-Bench 256K benchmark with a score of 60.3.

Cost and Efficiency

The cost per task of M3.1 Flash is 0.15 dollars per million input tokens (competitive price, median: 0.24 dollars) and 0.60 dollars per million output tokens (moderate price, median: 0.87 dollars). MiniMax expresses the value in cost per task, not per token.

Evaluation Metrics

MiniMax M3.1 Flash achieves a score of 43 on the Artificial Analysis Intelligence index, placing it well above the average of comparable models (median: 13). During the evaluation of this index, it generated 180 million tokens, which is very high compared to the median of 100 million.

Performance in Specific Tasks

On average, the evaluation of MiniMax-M3.1-Flash on the Intelligence Index costs 0.22 dollars per task. With 93 tokens per second, MiniMax-M3.1-Flash is slower than average (114).

Industry Context

The launch of M3.1 Flash by MiniMax comes a few weeks after Clement Delangue, CEO of Hugging Face, stated that China is currently leading the AI race and could dominate the sector by the end of the year.

Model's Real-World Applications

On October 19, 2026, the Chinese artificial intelligence company MiniMax announced the launch of M3.1 Flash. The model shows improved performance in a wide range of performance tests, reaching the level of cutting-edge models in complex tasks involving agents (cowork), computer use, software development (coding), and cybersecurity.

Business Workflow Capabilities

It is among the top performers on Automation-Bench-AA and is capable of carrying out real business workflows through SaaS systems, solving end-to-end problems in the areas of sales, marketing, operations, customer service, finance, and human resources. With a score of 76.2 on BankerToolBench, it excels in complex real-world challenges, such as multi-step planning, maintaining consistency between documents and files, financial compliance, and chaining tool calls.