nerkyor/Qwen3.6-27B-DSV4Pro-Thinking-Distill

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

nerkyor/Qwen3.6-27B-DSV4Pro-Thinking-Distill is a 27 billion parameter Qwen3.6-27B Dense model, fine-tuned by nerkyor using LoRA to distill the reasoning and agentic behavior of DeepSeek-V4-Pro. This model excels at complex reasoning tasks, showing significant improvements in GPQA and agentic performance while maintaining knowledge levels. It features a native Multi-Token Prediction (MTP) head for 2.3-2.6x single-stream inference acceleration, making it suitable for edge and desktop reasoning applications.

Loading preview...

Overview

nerkyor/Qwen3.6-27B-DSV4Pro-Thinking-Distill is a 27 billion parameter model based on the Qwen3.6-27B Dense architecture. It was fine-tuned using LoRA to distill the reasoning style and agentic behavior of DeepSeek-V4-Pro, specifically focusing on its "thinking-on" capabilities. This distillation process aims to teach the model how to reason and converge rather than injecting new knowledge or increasing its raw capability ceiling.

Key Capabilities & Differentiators

  • Enhanced Reasoning: Achieves a +7.1 percentage point improvement on GPQA-Diamond-198 (80.81%) and a +13.13 percentage point improvement on GPQA-Diamond-198 (81.82%) under streaming harness, demonstrating superior hard reasoning compared to its base model.
  • Improved Convergence: Significantly reduces unconverged answers, with GPQA finish=length cases dropping from 12 to 0, indicating the model "learns to converge" and provide complete responses.
  • Agentic Behavior: Shows improved performance on Agentic SOLO tasks (16/20 vs. 13/20 for the base), reflecting successful distillation of tool-calling and multi-step reasoning.
  • Multi-Token Prediction (MTP): Integrates a native MTP head, providing a 2.3-2.6x single-stream inference speedup across various quantization tiers (e.g., 26.8 TPS for Q4_K_M).
  • Knowledge Retention: Maintains or slightly improves MMLU scores (+0.2pp to +0.4pp), indicating that reasoning improvements do not come at the cost of general knowledge.
  • Robust Coding: Performance on coding-100 tasks is maintained or slightly improved (86/100 vs. 83/100).

Limitations

  • Distills Thinking Style, Not Capability: The model learns how to reason and converge, but black-box SFT does not inherently raise its knowledge ceiling.
  • Simulated Tool Execution: Tool execution results during training were simulated by a smaller model, not run in a real sandbox. This is an engineering trade-off for cost and speed, but it carries a "sim-to-real gap" risk, potentially leading to the model fabricating tool return values. Future versions plan to use real sandbox execution and rejection sampling to mitigate this.

Good For

  • Edge and Desktop Reasoning: Recommended as a local reasoning model for applications like Lynn Agent, especially the GGUF versions.
  • Complex Problem Solving: Ideal for tasks requiring multi-step reasoning, logical deduction, and agentic planning.
  • Applications Requiring Fast Inference: Benefits from the native MTP head for accelerated single-stream generation.

Finetunes
4 models