nbeerbower/Wichtel-Qwen3.6-27B

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 31, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Wichtel-Qwen3.6-27B by nbeerbower is a 27 billion parameter Qwen3.6-based model with a 32768-token context length, specifically optimized for advanced tool-use and code generation in the Hemlock language. It achieves a perfect 47/47 score on a tool-use benchmark and 67.6% on the Hemlock code generation benchmark. This model uniquely integrates multiple capabilities, including delegation and direct Hemlock code writing, without performance degradation.

Loading preview...

Wichtel-Qwen3.6-27B: A Specialized Tool-Use and Code Generation Model

Wichtel-Qwen3.6-27B is a 27 billion parameter model built on the Qwen3.6 architecture, developed by nbeerbower. It stands out for its exceptional performance in tool-use and its ability to generate code in the specialized Hemlock language. The model integrates four distinct adapters, allowing it to handle complex tasks like delegating engineering work to a code agent while also directly writing Hemlock code.

Key Capabilities

  • Advanced Tool-Use: Achieves a perfect 47/47 score on a comprehensive tool-use benchmark, demonstrating robust delegation, restraint (knowing when not to use a tool), inspection, execution, and memory management. It exhibits zero hallucinated tool names and deterministic behavior with greedy decoding.
  • Hemlock Code Generation: Scores 67.6% on the hembench benchmark, where generated programs are executed and their output verified against expected results. This includes handling syntax, standard library usage, algorithms, and systems-level tasks.
  • Multi-Capability Integration: Uniquely combines tool delegation and direct code generation for a specific language (Hemlock) without measurable interference between these capabilities, a significant achievement given the potential for conflict.
  • Robust Lineage: Developed through a deep merge of multiple adapters, including those for censorship circumvention (Bubba), delegation (egirl LoRA), and Hemlock SFT (Schierling).

Good For

  • Automated Agent Development: Ideal for scenarios requiring an LLM to interact with external tools and delegate tasks effectively, especially within the egirl tool schema.
  • Hemlock Language Development: Developers working with the Hemlock programming language can leverage this model for code generation and problem-solving.
  • Complex Task Automation: Use cases demanding a model that can both understand when to use a tool and when to directly generate specialized code.

Limitations

  • The 47-case tool-use benchmark is saturated, meaning it cannot differentiate between this model and potentially superior ones.
  • Weakness in Hemlock's standard library axis (2/5), requiring careful verification of API calls.
  • Supports single-turn tool use; it does not inherently handle follow-up actions or error recovery from tool calls.
  • Delegation is specifically bound to egirl's tool schema, requiring new training data for different tool sets.
  • Retains Chinese-language censorship gaps from its base model.

Finetunes
1 models