KAT-Coder-Pro V2.5 - Kuaishou's Agentic Coding Model

What is KAT-Coder-Pro V2.5

KAT-Coder-Pro V2.5 is the flagship Agentic Coding model released by Kuaishou KwaiKAT, achieving breakthroughs in three areas: long-horizon engineering capabilities, general Agentic capabilities, and large-scale Agentic reinforcement learning. Through its self-developed AutoBuilder, it increased the success rate of building real repository environments from 16.5% to 57.2%, accumulating over 100,000 runnable repositories; it also built the KwaiClawEnv general Agentic training system, combining multi-framework RL training with MOPD multi-expert fusion technology. The model scored 65.2 on SWE-Bench Pro and 94.2 on PinchBench, and is now fully available on the StreamLake platform.

Main Features of KAT-Coder-Pro V2.5

• Long-horizon repository engineering: Understands natural language issues, locates problems across file-spanning codebases, modifies code in multiple places, and runs tests to verify, achieving end-to-end delivery.
• General Agentic workflows: Executes complex multi-turn tasks such as data analysis, cross-system integration, batch document processing, and report generation, supporting an average of more than 10 rounds of tool interaction.
• Terminal and frontend capabilities: Supports terminal command execution and aesthetic frontend page generation, with no need to switch between multiple models.
• Multi-framework adaptation: Compatible with multiple Agent interaction protocols, with general problem-solving capabilities across tools and platforms.

Technical Principles of KAT-Coder-Pro V2.5

• AutoBuilder environment construction: The self-developed automated pipeline lets the model play the role of an "environment setup engineer," analyzing repository structure, generating configuration scripts, and verifying test execution in isolated sandboxes, increasing the success rate of building real repository environments from 16.5% to 57.2%, and accumulating over 100,000 runnable, verifiable environments covering 12 programming languages.
• Data flywheel and utilization of failed trajectories: Through full-process behavior filtering, it identifies high-value failed trajectories that were "one step short of success," and with targeted prompts re-executes them, converting about 20% of failed attempts into complete, reproducible, verifiable training data, enabling the model to learn error correction and recovery.
• KwaiClawEnv general Agentic training: The training environment is built in three layers: the Service layer dynamically expands the tool pool, converting open-source Skills into deployable services; the Task layer uses real business as seeds and derives massive task variants through tool-chain length and difficulty parameters; the Eval layer applies dual filtering by hard rules and model review, retaining only high-quality trajectories that are executable and behave naturally.
• Multi-framework reinforcement learning: Trains in differentiated Agent frameworks such as mini-swe-agent, Claude Code, Codex, and OpenClaw, enabling the model to learn "how to solve tasks."
• Asymmetric PPO and hierarchical rewards: Adopts an asymmetric PPO architecture, where during execution the model only sees real environment information; designs a three-layer reward mechanism—core task score, standard behavior constraints, and failed-trajectory incentives—encouraging exploration while correcting non-standard behavior.

How to Use KAT-Coder-Pro V2.5

• Apply for API access: Visit the StreamLake platform's KAT-Coder product page to submit an API KEY application.
• Review the integration documentation: Read the official development tool integration guide to learn about interface specifications and authentication methods.
• Integrate into development environments: Connect the API to IDEs such as VS Code and JetBrains, or to your own CI/CD pipeline.
• Submit engineering tasks: Describe the issue or workflow requirement in natural language to trigger the model to automatically analyze, modify, and verify in the repository.
• Acceptance and Iteration: Review the patches, test reports, and deliverables generated by the model, and optimize the task description based on feedback to obtain more precise results.

Core Advantages of KAT-Coder-Pro V2.5

• End-to-End Engineering Loop: The model can generate code snippets, and more importantly, complete the full long-horizon task from a vague issue to passing tests.
• Real-Environment Training: AutoBuilder increases the repository environment build success rate to 57.2%, and the model is battle-tested in over 100,000 real repositories.
• Stable in Complex Workflows: It scores 94.2 points on the PinchBench evaluation, which averages more than 10 rounds of tool interaction, with stable performance throughout the entire process.
• Framework Agnosticism: Through multi-framework RL training, the model does not depend on a specific interaction protocol and possesses strong generalization and transfer capabilities.
• Lossless Capability Expansion: The MOPD fusion mechanism ensures that when new long-horizon engineering and Agentic capabilities are added, existing capabilities such as front-end aesthetics are not sacrificed.

Project Address of KAT-Coder-Pro V2.5

• Project Official Website：https://streamlake.com/product/kat-coder
• arXiv Technical Paper：https://arxiv.org/pdf/2607.05471

Comparison of KAT-Coder-Pro V2.5 with Similar Competing Products

Application Scenarios of KAT-Coder-Pro

• Automated Bug Fixing: After receiving a GitHub issue, it automatically locates the root cause in a large code repository, generates a patch, and runs regression tests to completion.
• Data Analysis Report Generation: Reads raw data from multiple platforms and automatically cleans, aggregates, analyzes, and generates a Markdown business brief containing charts.
• Cross-System Data Integration: Calls multiple internal APIs and databases to automatically complete data extraction, transformation, loading (ETL), and exception handling.
• Batch Document Processing: Performs batch summarization, format conversion, key information extraction, and structured archiving for large numbers of technical documents or contracts.
• Rapid Front-End Prototype Generation: Based on product requirement descriptions, automatically generates HTML/CSS pages that conform to design specifications and supports iterative tuning.