KAT Coder

KAT Coder is Kuaishou's flagship AI coding model developed by Kwaipilot, specifically designed for advanced agentic code generation. Unlike traditional coding assistants, KAT Coder uses a multi-stage training pipeline including reinforcement learning, enabling it to autonomously complete complex software engineering tasks. With a 73.4% score on SWE-Bench Verified, it ranks among the best code generation models globally. Its key differentiators include: MoE architecture with ~72B active parameters, tool-use capabilities, long context support (128K tokens), and training on real Git commits and PRs.

KAT Coder itself is a closed-source, proprietary model accessible only through Kuaishou's StreamLake API. However, Kwaipilot has released open-source variants of the KAT series: KAT-Dev-32B and KAT-Dev-72B-Exp, both available on HuggingFace under Apache-2.0 license. These open models share the same training innovations as KAT Coder and achieve excellent performance (62.4% and 74.6% on SWE-Bench respectively), making them suitable for research and commercial use.

There are three main ways to access KAT models: (1) Open-source models: Download KAT-Dev-32B or KAT-Dev-72B-Exp from HuggingFace and run locally or on your infrastructure. (2) StreamLake API: Register at console.streamlake.com/wanqing and request API access to the flagship KAT Coder model. (3) Third-party platforms: Use services like Novita AI that provide KAT model playgrounds and APIs. The official documentation includes integration guides for Claude Code, vLLM, and various IDEs.

KAT Coder supports a wide range of programming languages and excels at 8 core task types: feature implementation, bug fixing, code refactoring, test generation, code review, documentation, deployment configuration, and interactive debugging. It's trained across 8 programming domains including ML/AI, web development (UI/UX), databases, security, systems programming, and more. The model handles multi-file reasoning, understands project context up to 128K tokens, and can orchestrate development tools like debuggers and code executors.

KAT Coder (73.4% SWE-Bench) rivals top closed-source models and outperforms most open-source alternatives. While Claude Sonnet 4.5 leads slightly (~75%), KAT Coder surpasses DeepSeek-Coder-33B (~58%) and Code Llama 34B (~55%). Unlike general-purpose models (GPT-4, Claude), KAT is specifically optimized for software engineering workflows with agentic capabilities, tool integration, and training on real development data (Git commits, PRs). The open KAT-Dev-72B-Exp achieves 74.6%, making it the top-performing open-source code model.

KAT Coder uses a Mixture-of-Experts (MoE) architecture based on Kuaishou's Qwen model family. It features approximately 72 billion active parameters at inference (out of over 1 trillion total parameters), providing massive capacity with computational efficiency. The model follows a four-stage training pipeline: (1) Mid-Training on agentic skills and Git data, (2) Supervised Fine-Tuning across diverse tasks, (3) Reinforcement Fine-Tuning with teacher trajectories, and (4) Large-scale Agentic RL with innovations like prefix caching, entropy pruning, and the SeamlessFlow distributed system.

Yes, KAT Coder is designed for production use. Kuaishou internally deploys KAT models in their Kwaipilot coding assistant, improving real-world development workflows. For enterprise users, the flagship KAT Coder is available via StreamLake API with proper SLAs and support. The open-source KAT-Dev models can be self-hosted for organizations requiring full control. Documentation covers integration with CI/CD pipelines, IDE plugins, and deployment configurations. The model's high accuracy (73.4% SWE-Bench) and tool-orchestration capabilities make it suitable for automating code reviews, bug fixes, test generation, and feature implementation in production codebases.

For the open-source models: KAT-Dev-32B requires approximately 65GB VRAM in FP16 (can run on 2×A100 40GB or 1×A100 80GB). KAT-Dev-72B-Exp needs ~145GB VRAM in FP16 (typically 2×A100 80GB or 4×A100 40GB). Quantized versions (FP8, INT4) are available on HuggingFace to reduce memory requirements. For production use, the closed-source KAT Coder is accessed via StreamLake API, eliminating infrastructure concerns. The API handles scaling and optimization, making enterprise-grade code generation accessible without dedicated GPU clusters.

Kwaipilot has an active roadmap for KAT Coder development. Recent updates include the release of KAT-Dev-32B (Sep 2025) and KAT-Dev-72B-Exp (Oct 2025). The team is training larger models, including a 200B-parameter MoE model with 40B active parameters. A comprehensive technical report on KAT Coder's training methodology is forthcoming. Future plans include enhanced tool ecosystems, richer IDE plugins, more specialized variants, and continued scaling of RL training. The team regularly publishes model weights, quantized versions, and improvements to the HuggingFace repository.

Official resources for KAT Coder include: (1) HuggingFace: Model cards, README files, and quickstart examples at huggingface.co/Kwaipilot. (2) Kwaipilot Blog: Technical posts and announcements at kwaipilot.github.io/KAT-Coder. (3) arXiv Papers: The KAT-V1 Technical Report and upcoming publications detail training methods. (4) StreamLake Documentation: Integration guides at console.streamlake.com/wanqing. (5) Community: HuggingFace Spaces for demos, GitHub discussions, and third-party tutorials on platforms like Novita AI. The community actively shares integration examples, deployment scripts, and use cases.