Kwaipilot Releases KAT-Coder-V2.5-Dev Open-Weight Model for Agentic Coding
Kwaipilot has released KAT-Coder-V2.5-Dev, an open-weight Mixture-of-Experts (MoE) model designed for agentic coding. The model features 35 billion total parameters, with only 3 billion active parameters during inference. This release provides the developer community with a highly efficient, open-weight model that claims state-of-the-art performance in its size class. By utilizing a Mixture-of-Experts architecture, it allows for advanced coding capabilities without requiring massive computational resources. Through Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), the model significantly reduced abnormal behaviors, reducing abnormal tool labels from 9.34% to 0.28% and eliminating single-turn repetitions.
## BACKGROUND
Agentic coding refers to the use of AI agents to assist in software development tasks like code generation, debugging, and testing. A Mixture of Experts (MoE) architecture divides a model into specialized subnetworks, routing inputs only to relevant "experts" to scale parameter capacity while keeping computational costs low.