Back
RCreddit.com
16
·1 days ago·Dev community · RSS

Qwen3.8-Next streaming - 150tps prefill, 3.6 tps decode on M5 Air

View original
Model release

Heat trend

Collecting trend data

The percentage is based on available heat signal, not comment count or independent people.

AI summary

A developer successfully adapted their DSv4 streaming stack to run Qwen3.8-Next on a 32GB M5 Air, achieving faster performance than a dense 27b model. Using a 3-bit version of the MoE, Qwen3.8-Flash-Next-MLX-oQ3-MTP, it reached 150 tps prefill and 3.6 tps decode for a 2k token prompt in low power mode. This compares favorably to the 27b-4bit model's 70 tps prefill and 3 tps decode.

Out of curiosity, I thought I'd see if I could adapt my DSv4 streaming stack from a few weeks ago to take Qwen3.8-Next. It worked, better than I thought - it actually runs faster on my 32GB M5 than the dense 27b does (admittedly not apples to apples as I decided to use a 3bit of the MoE, Qwen3.8-Flash-Next-MLX-oQ3-MTP, and the dense was 4bit).

For a 2k token prompt, running on low power mode on my M5, 3.8-Next-3bit gets 150 tps prefill, 3.6 tps decode. 27b-4bit gets 70tps prefill, 3 tps decode

Qwen3.8-Next streaming - 150tps prefill, 3.6 tps decode on M5 Air · BuzzRadr