Writing

Notes with the numbers

Write-ups of the kernel and inference work, with the charts and the commands to reproduce them. Most of them started as posts on r/LocalLLaMA. Feed

4 Oct 202620 min readMetal, ds4, MoE, benchmarks

Qwen3.8-Flash-Next on a 2021 M1 Max: 44 tok/s, and 35 tok/s with 400K tokens in the context

A 126B mixture-of-experts model on a 64 GB laptop, with kernels written for the M1, and how little the speed falls as the context grows.
26 Sep 20269 min readMetal, Splash, benchmarks

Splash on M1, part 2: 35B-A3B at 144 tok/s and a head-to-head with oMLX and MTPLX

Attention and the MoE expert layers move off MPP too. Speed, GPU temperature, energy per token and memory against two MLX engines on the same Mac.
24 Sep 20267 min readMetal, Splash, M1

Splash on a 2021 M1 Max: Qwen3.8-27B at 39 tok/s with custom Metal kernels

Splash only runs on M3 and newer. I ported it to my M1 Max and rewrote the kernels that depended on bfloat16 and Apple's MPP library, which doubled decode speed.