KyleHessling1/Qwopus3.6-27B-Fusion-BF16

Model Overview

Qwopus3.6-27B-Fusion-BF16 is a 27 billion parameter model built on the qwen35 hybrid architecture, developed by Kyle Hessling. It represents a novel merge of two specialized finetunes: Qwopus3.6-27B-v2 (reasoning) and Qwopus3.6-27B-Coder (code-execution). The core innovation lies in its layer-weighted delta-scale merge strategy, which protects early-layer reasoning representations while injecting coding skill in later layers, addressing instability issues found in previous merge attempts. This model is provided in full-precision bf16 safetensors format and supports a native context of 262,144 tokens, verified clean up to 60K.

Key Capabilities

• Fused Reasoning and Coding: Combines strong multi-step reasoning with robust code generation and debugging. Achieves 94.5% on HumanEval and 95.0% on GSM8K.
• Agentic Programming: Designed to function as a single, closed-loop agent for creative programming, capable of planning, writing, and self-debugging code (e.g., demonstrated by building interactive games and Kanban apps).
• Enhanced Stability: Exhibits improved temperature stability compared to its parent models, performing cleanly across a wider range of temperatures (0.2 to 0.9).
• Vision Encoder: Retains the base model's vision encoder, allowing for image input, though its multimodal quality is solid but uncharacterized.

Good For

• Single-agent creative programming: Ideal for tasks requiring an AI to independently plan, write, and debug code from a high-level goal.
• Complex problem-solving: Benefits from its strong reasoning capabilities combined with code execution for multi-step challenges.
• Developers seeking a versatile 27B model: Offers a balanced blend of reasoning and coding without the need for separate specialized models.