Swift 1.5 Qwen3.8 27b (A must-have for low thinking!)

A new 27B model, Swift 1.5 Qwen3.8 27b, has been released on Hugging Face. This model reportedly fixed a problem in 6 minutes on an old 3090 GPU, achieving 67 t/s. The developer expressed surprise that an IQ4_XS quant of a 27B model could outperform a major cloud model, especially for tasks requiring "low thinking."

This report highlights a 27B model's unexpected ability to outperform a major cloud model in "low thinking" tasks, unlike typical expectations for smaller, quantized models.