ManniX-ITA/Qwen3.6-27B-Omnimerge-v4

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ManniX-ITA/Qwen3.6-27B-Omnimerge-v4 is a 27 billion parameter language model based on the Qwen3.6 architecture, created by ManniX-ITA. It is a DARE-TIES merge of the Qwen3.6 base with three fine-tuned models, featuring a unique MLP-passthrough surgery to enhance reasoning stability. This model excels in complex reasoning tasks, achieving 78.28% on GPQA Diamond, and demonstrates strong performance in coding benchmarks like HumanEval and MBPP.

Loading preview...

Overview

ManniX-ITA/Qwen3.6-27B-Omnimerge-v4 is a 27 billion parameter model built upon the Qwen3.6 base, developed by ManniX-ITA. It utilizes a DARE-TIES (Omnimerge_v2 method) merge of the base model with three specialized Qwen3.6 fine-tunes. A key innovation is the MLP-passthrough surgery, which addresses a fragility in Qwen3.6's reasoning-tag emission policy, ensuring stable and reliable output.

Key Capabilities

  • Enhanced Reasoning: Achieves a notable 78.28% pass@1 on GPQA Diamond (full canonical greedy result), demonstrating significant improvement over its predecessor and base model in complex reasoning tasks.
  • Strong Coding Performance: Scores 83.54% on HumanEval pass@1 and 73.00% on MBPP pass@1 (corrected score), showcasing robust capabilities in code generation and problem-solving.
  • Multimodal Support: Preserves the vision tower from the Qwen3.6 base, enabling multimodal (vision-language) applications, particularly with MLX 4-bit VL quantizations for Apple Silicon.
  • Tool-Calling Proficiency: Scores 146.2 points on the tool-eval-bench hardmode, performing well in tool selection, parameter precision, and structured output.

Unique Differentiators

  • MLP-passthrough Surgery: This model specifically incorporates MLP layers verbatim from the clean Qwen3.6 base, a critical intervention to prevent issues with unclosed <think> tags and maintain high performance, especially in coding contexts.
  • Optimized for Reasoning Stability: The MLP-passthrough ensures the model's reasoning policy remains robust against small perturbations that typically destabilize Qwen3.6-based merges.
  • Efficient Inference: Available in various quantizations including GGUF for llama.cpp and MLX 4-bit for Apple Silicon, with a companion MTP version offering 2x decode speedup for interactive workloads.

Finetunes
5 models