Qwen3.6-35B-A3B JANG lands on Apple Silicon

This is more than another quant pack: it’s a practical argument for layer-aware compression on MoE models, where preserving attention precision matters a lot more than squeezing every last weight uniformly.

• –The 15 profiles give Mac users a real memory/performance ladder instead of a single compromise build
• –Activation-aware calibration plus MSE-all optimization points to quality-first quantization, not just size-chasing
• –MoE models are especially sensitive to naive quantization, so JANG’s higher-precision attention treatment is the key technical bet
• –Native support in vmlx and MLX Studio lowers the friction for local deployment; oMLX support broadens the ecosystem if the patch lands
• –The release also sets up a clear follow-on: if Qwen3-Coder-Next gets the same treatment, local coding workflows could benefit a lot