Huawei Ascend Enables MiniMax M2.7 Model Self-Evolution

Reports indicate that on the very day MiniMax open-sourced its self-iterating model MiniMax M2.7, Huawei Ascend AI foundation software and hardware announced zero-day compatibility. This allows developers to seamlessly deploy this flagship, self-evolving model on Huawei Ascend Atlas series products.
Key Breakthrough: The Model as Its Own "Researcher"
The core strength of MiniMax M2.7 lies in its advanced agentic capabilities:
Self-Iteration Cycle: The team guided an earlier model version into a research agent, enabling it to contribute to its own successor's development. It can autonomously construct complex agent harnesses, drive reinforcement learning, and optimize its memory, automating 30-50% of the workflow.
Programming and Office Proficiency: On the SWE-Pro programming benchmark, M2.7 scored 56.22%, performance on par with GPT-5.3-Codex. In professional office tasks, it leads open-source models with top GDPval-AA scores.
Exceptional Skill Adherence: Even in complex scenarios requiring over 2000 tokens, the model maintains a 97% instruction adherence rate.
Ascend Power: Hardware-Software Synergy Overcomes Compute Limits
To support the innovative architecture of M2.7, Huawei Ascend implemented deep optimizations across the stack:
Communication Acceleration: Introduced optimizations like ReduceScatter and AllGather for the model's FlashComm sequence parallelism, significantly boosting data transfer efficiency.
Operator Fusion: Deeply optimized full-chain Transformer Attention and large-scale MoE fusion operators, eliminating intermediate tensor read/write overhead.
Throughput Gains: Achieved adaptive load balancing for multi-data-parallel scenarios, drastically reducing prefill phase interference on decoding.
Real-World Applications: From Software Engineering to Interactive Entertainment
With the full-cycle inference deployment support from Huawei Ascend, MiniMax M2.7 proves its value across diverse fields:
Software Engineering: Excels in deep tasks like log analysis, bug localization, code refactoring, and Android development.
Interactive Entertainment: Through the OpenRoom interactive system, AI is embedded into web-based GUI spaces, enhancing character consistency and dialogue flow.
Professional Office Work: Leverages robust environmental interaction for highly complex productivity tasks.
Conclusion: Compute Infrastructure Dictates the Pace of Evolution
From Huawei Ascend's initial adaptation taking hours to today's zero-day synchronized support, domestic AI infrastructure provides the most solid foundation for top-tier models like MiniMax pursuing "self-evolution" through rapid response capability.
Related article
Qwen-Audio-3.0-TTS: Qwen's Large Voice Synthesis Model Cuts First Packet Delay to 300ms
The competition in voice technology among major AI developers has intensified with a new entry. On July 20, Qwen released its latest audio synthesis model, Qwen-Audio-3.0-TTS, shifting control from complex technical configurations to simple natural l
Claude’s Cowork feature now recalls your past chat instructions
Anthropic’s update to Claude’s memory system removes a major friction point in using AI agents: the repetitive need to re-explain context that the AI already retains. By merging the memory architectures of standard chat and Claude Cowork, Claude now
Chen Tianqiao Returns After Nine Years, Investing $2B in Discovery-Driven AI
From Gaming Mogul to Consciousness Pioneer: The Return of a Former BillionaireAfter nine years of retirement, Shanda Group founder Chen Tianqiao gave his first international media interview in March 2026. Once the youngest Chinese billionaire and gam
Related Special Topic Recommendations
Comments (0)
0/500

Reports indicate that on the very day MiniMax open-sourced its self-iterating model MiniMax M2.7, Huawei Ascend AI foundation software and hardware announced zero-day compatibility. This allows developers to seamlessly deploy this flagship, self-evolving model on Huawei Ascend Atlas series products.
Key Breakthrough: The Model as Its Own "Researcher"
The core strength of MiniMax M2.7 lies in its advanced agentic capabilities:
Self-Iteration Cycle: The team guided an earlier model version into a research agent, enabling it to contribute to its own successor's development. It can autonomously construct complex agent harnesses, drive reinforcement learning, and optimize its memory, automating 30-50% of the workflow.
Programming and Office Proficiency: On the SWE-Pro programming benchmark, M2.7 scored 56.22%, performance on par with GPT-5.3-Codex. In professional office tasks, it leads open-source models with top GDPval-AA scores.
Exceptional Skill Adherence: Even in complex scenarios requiring over 2000 tokens, the model maintains a 97% instruction adherence rate.
Ascend Power: Hardware-Software Synergy Overcomes Compute Limits
To support the innovative architecture of M2.7, Huawei Ascend implemented deep optimizations across the stack:
Communication Acceleration: Introduced optimizations like ReduceScatter and AllGather for the model's FlashComm sequence parallelism, significantly boosting data transfer efficiency.
Operator Fusion: Deeply optimized full-chain Transformer Attention and large-scale MoE fusion operators, eliminating intermediate tensor read/write overhead.
Throughput Gains: Achieved adaptive load balancing for multi-data-parallel scenarios, drastically reducing prefill phase interference on decoding.
Real-World Applications: From Software Engineering to Interactive Entertainment
With the full-cycle inference deployment support from Huawei Ascend, MiniMax M2.7 proves its value across diverse fields:
Software Engineering: Excels in deep tasks like log analysis, bug localization, code refactoring, and Android development.
Interactive Entertainment: Through the OpenRoom interactive system, AI is embedded into web-based GUI spaces, enhancing character consistency and dialogue flow.
Professional Office Work: Leverages robust environmental interaction for highly complex productivity tasks.
Conclusion: Compute Infrastructure Dictates the Pace of Evolution
From Huawei Ascend's initial adaptation taking hours to today's zero-day synchronized support, domestic AI infrastructure provides the most solid foundation for top-tier models like MiniMax pursuing "self-evolution" through rapid response capability.
Qwen-Audio-3.0-TTS: Qwen's Large Voice Synthesis Model Cuts First Packet Delay to 300ms
The competition in voice technology among major AI developers has intensified with a new entry. On July 20, Qwen released its latest audio synthesis model, Qwen-Audio-3.0-TTS, shifting control from complex technical configurations to simple natural l
Claude’s Cowork feature now recalls your past chat instructions
Anthropic’s update to Claude’s memory system removes a major friction point in using AI agents: the repetitive need to re-explain context that the AI already retains. By merging the memory architectures of standard chat and Claude Cowork, Claude now
Chen Tianqiao Returns After Nine Years, Investing $2B in Discovery-Driven AI
From Gaming Mogul to Consciousness Pioneer: The Return of a Former BillionaireAfter nine years of retirement, Shanda Group founder Chen Tianqiao gave his first international media interview in March 2026. Once the youngest Chinese billionaire and gam





Home






