schneewolflabs/Wichtelchen-Qwen3.5-9B

Model Overview

Wichtelchen-Qwen3.5-9B is a 9 billion parameter model developed by schneewolflabs, built upon the Qwen3.5-9B-Base architecture. It is a smaller, more efficient version of the Wichtel-Qwen3.6-27B, specifically fine-tuned for local operation and delegating engineering work to a code agent. The model demonstrates strong performance in code generation, achieving 56.1% on hembench for Hemlock code, which is close to the 27B model's 67.6% at a significantly reduced parameter count.

Key Capabilities

• Code Agent Delegation: Excels at handing off engineering tasks to a code agent, achieving 10/10 on the delegate axis of the egirl 47-case benchmark.
• Hemlock Code Generation: Proficient in generating Hemlock code, scoring 56.1% on the hembench weighted metric.
• Tool Use: Incorporates Hemlock tool use and is designed for single-turn tool interactions, bound to the egirl tool schema.
• Behavioral Consistency: Maintains general capabilities, with ARC scores and wiki perplexity remaining stable across its training rungs, indicating focused behavioral improvements without sacrificing core language understanding.

Limitations

• Call Formatting: May occasionally emit malformed tool calls, requiring a lenient parser or additional preference training.
• Restraint: Can sometimes over-delegate or reach for commands like git_status or execute_command when a direct answer is expected.
• Hemlock stdlib: Similar to other models in its family, it has a weaker grasp of standard library function names, necessitating documentation checks.

Use Cases

This model is particularly well-suited for applications requiring automated code generation, agent-based engineering task delegation, and scenarios where efficient local operation with strong tool-use capabilities is paramount.