AMAImedia/Qwen3.6-27B-Qwimi-Coder-MTP-NOESIS-BF16

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 1, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AMAImedia/Qwen3.6-27B-Qwimi-Coder-MTP-NOESIS-BF16 is a 27.8 billion parameter Qwen 3.6-based model, fine-tuned by trjxter and repackaged by AMAImedia, specifically optimized for coding assistance, tool-calling, and agentic behaviors. This text-only SFT, merged to full-precision BF16, excels at code generation, debugging, and refactoring, demonstrating significant task-level latency reductions. It features a hybrid architecture with GatedDeltaNet linear-attention layers and supports a 16,384-token context length for its specialized tasks. The model is intended for developers requiring efficient and accurate code-centric AI solutions.

Loading preview...

Overview

AMAImedia/Qwen3.6-27B-Qwimi-Coder-MTP-NOESIS-BF16 is a 27.8 billion parameter model based on Qwen 3.6, fine-tuned by trjxter and repackaged by AMAImedia. It is a text-only supervised fine-tune (SFT) merged to full-precision BF16, with its vision tower frozen. The model was trained on a mixed corpus, predominantly coding data (~82.5% of training tokens), alongside tool-calling and agentic (SWE-agent) data, to enhance its specialized capabilities while preserving coding proficiency. It supports a validated context length of 16,384 tokens.

Key Capabilities

  • Coding Assistance: Excels at code generation, debugging, refactoring, and explanation with visible chain-of-thought reasoning.
  • Tool-Calling: Features native function/tool calling using Qwen 3.6's XML-style format.
  • Agentic Behavior: Supports basic multi-turn tool use over a repository, inspired by SWE-agent.
  • Performance Efficiency: Achieves significant reductions in task-completion time, up to 49.4% faster on custom benchmarks, by generating more direct and concise responses.

Benchmarks & Performance

Evaluations against the base Qwen 3.6 27B model show a modest but measurable improvement in overall pass rates, particularly in custom coding (+4.00 pts) and HumanEval (+14.29 pts). While agentic accuracy saw a slight regression, the model consistently demonstrated substantial reductions in average wall time across various coding and tool-use tasks. The model's architecture includes GatedDeltaNet linear-attention layers, contributing to its performance.

Good For

  • Developers requiring a specialized model for code generation and manipulation.
  • Applications involving function/tool calling within a coding context.
  • Use cases demanding efficient task completion where response length and directness are critical.
  • Integration into automated development workflows or coding assistants.

Finetunes
1 models