cyberagent/CAT-Paws-8B

CAT-Paws-8B: Japanese Agentic LLM

CAT-Paws-8B, developed by CyberAgent, is an 8 billion parameter agentic large language model built upon the Qwen3-Swallow-v0.2 architecture. Its core distinction is its ability to reason in Japanese, even when the input is in English, making it highly suitable for Japanese-centric agentic applications.

Key Capabilities

• Multi-turn Agentic Interactions: Designed for complex, multi-turn dialogues, particularly in scenarios requiring sequential actions.
• Coding and Terminal Interaction: Proficient in tasks involving code generation, execution, and interaction with terminal environments.
• Tool Usage: Supports tool calling by accepting tool descriptions and formats via system prompts or user messages, with a recommended JSON interface for tool calls.
• Japanese Reasoning: Generates reasoning traces primarily in Japanese, a unique feature for non-English agentic workflows.

Performance and Limitations

While CAT-Paws-8B shows competitive performance against Qwen-3-8B in agentic benchmarks like j-tau-bench (especially in Japanese telecom tasks) and humaneval-fix (45.7% accuracy), it is noted to be a lightweight solution for simpler tasks compared to frontier models. For single-turn tasks, its sibling model, CAT-Thinking-8B, is recommended as CAT-Paws may generate longer, sometimes repetitive, responses and can struggle with conflicting instructions. The model is trained with a maximum output token length of 4096, though it may generate longer outputs.

Good for

• Japanese-speaking agents: Ideal for building agents that need to reason and interact in Japanese.
• Coding and automation: Use cases involving code generation, script execution, and automated terminal operations.
• Tool-augmented LLM applications: Scenarios where the LLM needs to interact with external tools and APIs, especially with a focus on Japanese language processing.