Skip to main content

AIGCode's L3 Programming Model Boosts Computing Power Utilization to 65%

The AI programming landscape is shifting. While many vendors are still battling over L2-level code completion, a two-year-old technical team called AIGCode is trying to jump ahead with a full-stack solution. Their core product, AutoCoder.cc, aims to move beyond the 'assisted driving' stage of L2 and into the 'autonomous driving' of L3.

AIGCode's key insight is that the shortcomings of large models often lie not in fine-tuning, but in the base model itself. Their goal is to create a system that can complete the entire development cycle—front-end, back-end, and database—based on natural language requirements, and even handle operations and deployment.

To achieve this, AIGCode has built a positive feedback loop: they produce high-quality samples through practical use of L3-level products, then use those samples to train stronger base models. This concept, which they call the '3No Paradigm'—no reliance on NVIDIA's single path, surpassing L2 assistance, and rejecting blind scaling—makes computing power utilization a key competitive barrier.

On the computing power front, AIGCode has taken an unusual path. Amid tight resources, the team achieved a 65% computing power utilization (MFU) for MoE architecture on domestic chip clusters through extreme software optimization and deep engineering practices. This achievement not only demonstrates their expertise in high-performance optimization but also proves that the domestic hardware ecosystem can bridge the gap with mainstream hardware when supported by advanced software.

Beyond the 'grunt work' of computing power, fundamental research is another major focus. The tree-structured position encoding (TPE), developed over 18 months, has shown better performance than the industry-standard RoPE approach in handling long texts and extrapolation capabilities. Innovations like TPA (lossless linear attention mechanism) and PLE (expert decoupling architecture) further enrich their technology matrix. These research achievements have received recognition from top academic conferences and demonstrated actual efficiency in commercial applications, enabling developers to build complex applications within minutes.

Key Points

  • AIGCode's L3-level programming model aims to automate the entire development cycle.
  • The company achieved 65% computing power utilization on domestic chips.
  • Innovations include tree-structured position encoding (TPE) and the '3No Paradigm'.