RT by @awnihannun: Pleased to publish another DWQ - this one of Qwen3.5-27B in 4bit on MLX - using Qwen's Int4 GPTQ quant as a base, quantizing attn + embedding... (opens in new tab)

<p>Pleased to publish another DWQ - this one of Qwen3.5-27B in 4bit on MLX - using Qwen's Int4 GPTQ quant as a base, quantizing attn + embedding params as well at 4bit32gs, and then DWQing.</p> <img src="http://twitter.macworks.dev/pic/media%2FHEI1llEbwAAubbC.jpg" style="max-width:250px;" />