According to reports, on the day when
Key Breakthroughs: Let the Model Become Its Own "Researcher"
The core competitiveness of
Self-Iteration Cycle: The development team guided an early version of M2 into a research-type agent, allowing the model to participate in the iteration of the next-generation model. It can autonomously build complex Agent Harnesses, drive reinforcement learning, and optimize its own memory, achieving 30-50% automation of workflow.
Programming and Office Excellence: In the programming test SWE-Pro, M2.7 achieved a 56.22% accuracy rate, matching GPT-5.3-Codex; in professional office scenarios, it ranked first among open-source models with a GDPval-AA score.
Exceptional Skill Adherence: Even in complex skill scenarios involving more than 2000 Tokens, the model maintains a 97% high adherence rate.
Ascend Power: Soft-Hardware Collaboration Eliminates Computing Bottlenecks
To match the innovative architecture of
Communication Acceleration: For the model's innovative FlashComm sequence splitting, ReduceScatter and AllGather communication optimizations were introduced, significantly improving data transmission efficiency.
Operator Fusion: Deeply optimized Transformer Attention full-chain fusion operators and MoE large-scale fusion operators, completely eliminating intermediate tensor read and write costs.
Throughput Performance Improvement: Achieving adaptive load balancing in multi DP concurrent scenarios, significantly reducing the interference of prefilling (prefill) on decoding (decode).
Practical Applications: From Software Engineering to Interactive Entertainment
With the full-cycle inference deployment support provided by
Software Engineering: Covering deep areas such as log analysis, bug location, code refactoring, and Android development.
Interactive Entertainment: Through the interactive system OpenRoom, AI interaction is embedded into the Web GUI space, enhancing character consistency and dialogue coherence.
Professional Office Work: Leveraging strong environmental interaction capabilities, it delivers highly complex productivity tasks.
Conclusion: Computing Infrastructure Determines Evolution Speed
From
