lzy510016411/qwen3.5-9b-distill-agent-instruct

Overview

Qwen3.5-9B-Distill-Agent-Instruct is a 9 billion parameter model, a distilled and post-trained version of the Qwen3.5-9B base model. It is specifically designed for agentic applications, learning from a curated mixture of multi-model agent trajectories. This model significantly enhances capabilities related to tool use, structured function calling, and multi-step agent execution, while also improving code-oriented tasks and general instruction following.

Key Capabilities

• Agent-oriented post-training: Trained on long-horizon interaction traces, including planning, tool calls, and recovery behavior.
• Enhanced tool-use decisions: Explicitly learns when to use a tool and when to provide a direct answer.
• Multiple and parallel function calling: Improves coordination for concurrent tool execution.
• Irrelevance handling: Better resistance to invoking tools that don't match the user's request, showing a +5.42 pp improvement on the BFCL Irrelevance category.
• Structured output: Trained on explicit function schemas for easier integration with agent runtimes.
• Code and instruction following: Includes cleaned code examples and general instruction trajectories.
• Long-context foundation: Inherits Qwen3.5-9B's native 262,144-token context window.
• Reduced Latency: Demonstrates a 14.2% decrease in mean latency compared to the base model in evaluation.

Good for

• Agent orchestration and development: Ideal for building applications requiring complex agentic behavior.
• Function calling and tool selection: Excels in scenarios where models need to make reliable tool-use decisions.
• Multi-step task execution: Suited for tasks that involve planning and executing multiple steps.
• Coding assistance: Benefits from cleaned code instruction data in its training mixture.
• Instruction following: Strong performance in general instruction-based tasks.