Domestic large model development has welcomed another significant product release. On June 1, the domestic large model company MiniMax (00100) officially launched its new generation general-purpose model, MiniMax M3. The M3 model employs a novel self-developed sparse attention architecture called MiniMax Sparse Attention (MSA), achieving generational breakthroughs in several key areas including programming and agent capabilities, ultra-long context, and native multimodal understanding. Reportedly, M3 is the first domestic large model to simultaneously possess the three core capabilities of "advanced coding ability, 1M ultra-long context, and native multimodality." It is also currently the only open-source option globally that offers this complete combination of capabilities.
Core Breakthrough: Rewriting the Underlying Attention Mechanism to Support 1M Context The integration of M3's three core capabilities is underpinned by its self-developed sparse attention architecture, MSA. Compared to traditional full attention mechanisms, MSA can significantly reduce computational costs in long-context scenarios and extend the context window to 1 million tokens. This means the model can retain a more complete information chain in a single inference when processing long documents, complex code repositories, and multi-round collaborative tasks. MiniMax disclosed that at the 1 million context scale, M3's computational cost per token is only about 1/20th of the previous generation model, with a significant improvement in inference efficiency. Beyond the model architecture upgrade, MiniMax has also further optimized the underlying inference operators. By redesigning data reading and computation paths, the relevant performance has improved by over four times compared to mainstream open-source solutions. Industry observers view this as an important new variable in the global large model competition. As the complexity of Agent tasks continues to increase, "longer context, more stable memory, and lower-cost inference" are becoming key capabilities determining product usability.
Programming and Agent Capabilities: Multiple Metrics Rival Top International Closed-Source Models M3 shows marked improvement in Coding & Agentic capabilities, achieving internationally leading levels in several authoritative international evaluations covering dimensions such as software engineering, terminal execution, efficiency, and protocol understanding. On the SWE-Bench Pro benchmark measuring coding ability, MiniMax M3 surpassed GPT-5.5 and Gemini 3.1 Pro, approaching the performance of Opus 4.7. On the SVG-Bench benchmark for comprehensive evaluation of SVG generation performance, MiniMax M3 exceeded Opus 4.7. On the multimodal test set OmniDocBench, MiniMax M3 scored higher than Gemini 3.1 Pro. On the end-to-end evaluation framework Claw-Eval for autonomous Agents, MiniMax M3 achieved the highest score. It is reported that M3 innovatively introduced an interactive user simulator framework during its programming and Agent training. By simulating the behavioral patterns of real developers in collaborative processes, the model is exposed to interaction scenarios closer to production environments during both training and evaluation phases. The industry believes that from code development and research analysis to cross-application collaborative execution, Coding & Agentic capabilities are gradually becoming a new competitive focus for top global models. MiniMax's emphasis on strengthening this capability is seen by the market as a strategic move to position itself for the next phase of AI product evolution.
Native Multimodality: Training Data Scale Pushed to 100 Trillion Token Level MiniMax stated that M3 was trained from the outset with mixed multimodal data including text, images, and videos, and has further expanded in data scale and training pipelines. The model not only supports image and video understanding but also possesses desktop operation capabilities, enabling it to perform Computer Use tasks in complex cross-application environments. M3 is a model trained with multimodal mixed data from the very beginning (Step 0). MiniMax emphasized in its report that Interleaved data—where text and other modalities like images are naturally arranged alternately in a sequence—contributes more critically to model performance than generally assumed. After reconstructing the entire data pipeline for such data, MiniMax has been able to scale the training data token volume to the level of 100 trillion. This signifies that the model's capabilities are extending further from language understanding into real digital environments. Whether for office automation, enterprise software operation, or more complex productivity scenarios, the speed at which AI is entering the practical execution layer is accelerating noticeably.
Co-Trained Product: MiniMax Code Receives Major Update On the same day, MiniMax Code also received an update. As an Agent product specifically designed for M3 and trained alongside it, MiniMax Code can fully leverage M3's capabilities in long context, Coding/Agentic tasks, and native multimodality, making it the preferred Agent to pair with MiniMax-M3. For long-range, complex tasks, MiniMax Code's Agent Team can decompose large tasks into multi-stage, concurrently executable, and dynamically adjustable workflows, which are then advanced collaboratively by a cluster of Agents. On the commercialization front, MiniMax simultaneously launched a Token Plan subscription scheme. The Plus version costs 49 yuan per month, providing 600 million tokens; the Max version costs 119 yuan per month, providing 1.8 billion tokens; and the Ultra version costs 469 yuan per month, providing 5.5 billion tokens. Industry insiders believe that with the release of M3, MiniMax's positioning in the global AI competition is becoming clearer. It is differentiating itself with a cutting-edge model characterized by "open-source + multi-capability integration," filling a gap in the domestic AI ecosystem in this dimension. The AI industry remains in a phase of rapid evolution. As model capabilities continue to approach real-world work scenarios, the next round of competition around Agents has already commenced ahead of schedule.
Comments