TeichAI/Qwen3.5-27B-Claude-Opus-4.6-Distill
TeichAI/Qwen3.5-27B-Claude-Opus-4.6-Distill is a 27 billion parameter language model based on the Qwen3.5 architecture, fine-tuned using LoRAs derived from Claude Opus 4.6 data. Developed by TeichAI with contributions from EclipseMist and crownelius, this model is designed for enhanced performance across coding, creative writing, visual understanding, and general-purpose tasks. It demonstrates improved benchmark scores over its base model in areas like ARC Challenge, HellaSwag, MMLU, TruthfulQA, and Winogrande, leveraging a 32768 token context length.
Loading preview...
Model Overview
TeichAI/Qwen3.5-27B-Claude-Opus-4.6-Distill is a 27 billion parameter model built upon the unsloth/Qwen3.5-27B base. It has been fine-tuned using LoRAs provided by EclipseMist, incorporating datasets such as crownelius/Opus-4.6-Reasoning-2100x-formatted and personal Claude data. This distillation process aims to imbue the model with capabilities akin to Claude Opus 4.6.
Key Capabilities and Performance
This model shows notable improvements over its base model across several benchmarks:
- ARC Challenge: Achieves 0.461, surpassing the base model's 0.435.
- HellaSwag: Scores 0.613, an increase from the base model's 0.574.
- MMLU: Reaches 0.233, slightly higher than the base model's 0.230.
- TruthfulQA_MC2: Scores 0.610, compared to the base model's 0.599.
- Winogrande: Achieves 0.769, outperforming the base model's 0.749.
It maintains a competitive score on GPQA Diamond Zero-shot at 0.283.
Recommended Use Cases
This model is versatile and recommended for a range of applications, including:
- Coding: Generating and assisting with code.
- Creative Writing: Producing various forms of creative text.
- Visual Understanding: Processing and interpreting visual information (as a native-multimodal base model).
- General Purpose: Handling a broad spectrum of natural language tasks.