trjxter/Qwimi-3.6-27B-Coder-MTP-BF16

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Qwimi-3.6-27B-Coder-MTP-BF16 by trjxter is a 27.8 billion parameter, coding-focused supervised fine-tune of the Qwen 3.6 27B model, merged to full-precision BF16. It is specifically optimized for coding assistance, tool-calling, and agentic behaviors, with a training focus on code generation, debugging, and refactoring. This model excels in coding tasks and structured tool usage, demonstrating significant reductions in task completion time compared to its base model.

Loading preview...

Model Overview

trjxter/Qwimi-3.6-27B-Coder-MTP-BF16 is a 27.8 billion parameter model, fine-tuned from unsloth/Qwen3.6-27B. It features a hybrid architecture combining standard attention layers with GatedDeltaNet linear-attention layers. This supervised fine-tune (SFT) is text-only, with the base model's vision tower frozen and untouched during training. The model was trained with a maximum sequence length of 16,384 tokens and is released as a merged, full-precision BF16 checkpoint.

Key Capabilities

  • Coding Assistance: Optimized for code generation, debugging, refactoring, and explanation with visible chain-of-thought reasoning.
  • Tool-Calling: Supports native function/tool calling using Qwen 3.6's XML-style format.
  • Agentic Behavior: Capable of basic agentic tasks, including multi-turn tool use over a repository (SWE-agent style).
  • Performance: Achieves a small accuracy improvement and a significant reduction in task-completion time (up to 49.4% faster) on coding and tool-use benchmarks compared to the base Qwen 3.6 27B.

Intended Use Cases

  • Software Development: Ideal for developers seeking assistance with various coding tasks.
  • Automated Workflows: Suitable for applications requiring function calling and basic agentic operations.

Limitations

  • Text-Only: Not validated for image/vision input.
  • Agentic Data Thinness: Agentic capabilities, while present, are based on a comparatively smaller and shorter dataset, leading to some regression in agentic accuracy on custom benchmarks.
  • Context Length: Validated up to 16,384 tokens; behavior beyond this length is unvalidated.

Finetunes
1 models