Qwen3.5-9B-Kimi-k3-Distilled
Model Summary
Qwen3.5-9B-Kimi-k3-Distilled is a fine-tuned, reasoning-focused model derived from the Qwen architecture via knowledge distillation from Kimi K3 (moonshotai/kimi-k3).
Unlike standard instruction-tuned models that heavily rely on speculative internal chain-of-thought hallucination, this model is explicitly optimized for environment-grounded agentic workflows. It natively integrates reasoning with tool execution, prioritizing direct environment interaction (e.g., executing shell commands, listing files, reading logs) over ungrounded hypothetical reasoning loops.
Key Features & Behavior
- Grounded Tool Calling First: Prefers inspecting files, running test scripts, and querying environments over guessing code structure or hallucinating codebase state.
- Behaviour-Preserving Distillation: Inherits the multi-step trajectory planning, tool selection, and state tracking behaviors of Kimi K3.
- Loop Resilience: Specialized in build-test-fix loops and verification-driven task completion (e.g., SWE-bench style debugging, Bash/CLI automation).
- Parallel & Structured Tool Calls: Supports structured tool schemas and parallel tool execution when required by the agent harness.
Intended Use
This model is designed primarily for:
- Supervised Fine-Tuning (SFT) & Distillation Baselines: Serves as a strong base or control model for agentic datasets.
- Coding & Shell Automation Agents: Executing bash commands, repository exploration, bug fixing, and continuous integration troubleshooting.
- Multi-Step Agentic Trajectories: Analysis and deployment of tasks requiring long-horizon planning, parallel calls, tool selection, state tracking, and self-verification.
Out-of-Scope / Misuse
- Direct deployment without sandbox isolation (the model is trained to execute shell commands and code).
- General ungrounded long-form creative writing where strict environment feedback is absent.
Training Details
Dataset & Trajectory Source
The dataset consists of behaviour-preserving instruction-following, tool-use, and agent trajectories distilled from Kimi K3 (moonshotai/kimi-k3).
Programming Languages Covered
The training trajectory drop includes multi-language code contexts, scripts, and tool calls across:
Python, TypeScript, English, Go, Bash, C, C#, Java, C++, Ruby, Assembly, PowerShell, Rust, and general technical plain text (unknown).
Task Distribution
The training trajectory mix emphasizes structural tool integration rather than static QA:
- Environment exploration and file inspection (
ls,grep,cat, repo navigation). - Iterative error handling and test-driven fixes.
- Multi-tool chaining and tool argument serialization.
Benchmark & Behavioral Comparison
During internal evaluations comparing Qwen3.5-9B against Qwen3.5-9B-Kimi-k3-Distilled on agentic/coding benchmarks:
| Model | Grounded Tool Call Rate | Speculative Hallucination Rate | Task Completion Efficiency |
|---|---|---|---|
| Qwen3.5-9B (Base) | Low (tends to guess) | High (long hypothetical text chains) | Sub-optimal in agent loops |
| Qwen3.5-9B-Kimi-k3-Distilled | High (Immediate Execution) | Low (Action-driven reasoning) | High (Fast convergence) |
Example Trajectory Difference
- Task: Debug a broken deployment script (
release.sh) in a workspace. - Base Model (Qwen3.5-9B): Generates long speculative reasoning loops trying to guess missing code and hypothetical argument-passing bugs without checking the actual directory.
- Distilled Model (Qwen3.5-9B-Kimi-k3-Distilled): Immediately triggers a
bashtool call (ls -la/cat release.sh) to inspect the real file before drawing conclusions.