/NukeTorch.AI

Qwen3.5-9B fine-tuning on four MacBook Pros

Apple’s WWDC26 LoRA fine-tuning recipe, pipelined across four MacBook Pro M5 Max with NukeTorch.

5,358.5
tokens/s · NukeTorch, four MacBook Pro M5 Max
~600
tokens/s · Apple’s WWDC26 run, four M3 Ultra (Apple’s figure)
8×+
faster than Apple’s WWDC26 run, every run

One Mac and four Macs both run 1,000 steps of Apple’s recipe. On one Mac it trains 4 examples per step, Apple’s one-Mac setting; on four Macs, 16 per step (4 per Mac). The four-Mac job does not fit in one 48 GB Mac. Speeds are compared in tokens per second.

Read the reportFull WikiSQL reportOne-Mac report

Qwen3.8 Flash Next on four MacBook Pros

A 4-bit model of 106.3 GB, run across four 48 GB MacBook Pro M5 Max with NukeTorch pipeline parallelism.

177.7
tokens/s · one user
859.0
tokens/s · 32 users at once

For reference, 4-bit vs 4-bit: one NVIDIA H100 NVL runs one user at 151 tokens/s (published). NukeTorch’s one user on four MacBook Pros ran 1.18× that.

Read the report

NVIDIA’s GR00T robot model on four MacBook Pros

NVIDIA’s GR00T N1.7 robot policy, 192 clients at once on four MacBook Pro M5 Max with NukeTorch.

152.07
policy chunks/s, four MacBook Pro M5 Max
2.45×
faster than the fastest published H100 result
4.24×
faster than NVIDIA’s own published H100 figure
Read the report