An Experimental Kernel Lifts Qwen3.8-Flash-Next to 150 Tok/s on a Mac

Same Mac, same model, a new kernel: about 40% faster on prose, by his numbers.

@ashxhart says Qwen 3.8 Next Flash at 4 bits now hits over 150 tok/s on prose and 230 on code on his M3U Mac Studio, up from around 110. He calls it an experimental kernel he has been working on. @MiaAI_lab shared it: “150 tokens per sec on prose with Qwen3.8-Flash-Next.”

Separately, @ashxhart shipped TensorFold 1.1.3. It fixes image-chat caching for Flash Next, tool calls, reasoning and checkpoint loading, and restores Windows builds.