ModelBest Shifts AI Development Toward Automated Self-Improvement
Recursive self-improvement is moving from theoretical research into production, as Beijing-based ModelBest demonstrates AI systems capable of building and optimizing their own software. By removing human engineers from the loop, the firm has achieved training speeds exceeding NVIDIA’s Megatron and automated complex stencil optimization in record time.
The company’s recent trajectory centers on Forge Engineering, a paradigm that replaces traditional human-led software development with autonomous coding agents. In May, ModelBest teamed with the OpenBMB community to launch ForgeTrain, a framework that successfully trained the MiniCPM5-1B model. This system outperformed NVIDIA’s Megatron in training speed, marking a shift toward AI-designed infrastructure.
This momentum continued through the summer with two distinct technical milestones. During the World Artificial Intelligence Conference in July, an agent powered by GLM-5.2 iteratively optimized a training framework in 18 hours to build the MiniCPM5-130M model from scratch, matching the performance of Google's Gemma 3 270M. By August, the company introduced ForgeStencil, an automated system that optimized over 100 stencils for scientific and industrial software within a single week.
These results signal a fundamental move away from general-purpose frameworks, which have long been hampered by the high cost of human programming. By leveraging AI to craft customized software, ModelBest aims to drive down compute costs and accelerate the development cycle. This approach suggests that the next generation of artificial intelligence may be defined by systems that refine their own architecture, making advanced technology accessible to a wider array of users.
Comments (0)
No comments yet. Be the first!