Tensorunfolded
1def unfold(T: nuke.Tensor) -> nuke.Tensor:
2 """Values are fixed; structure is a choice.
3 m ≤ Φ ≤ c·m — unfolding moves the store to Φ = m.
4 m: number of draws · Φ: fetch count · c: word capacity"""Neural Unfolded Kernel Engine
NukeTorch is a Tensor Engine. Its kernels and framework are built on our own mathematical formulation, with three principles:
Unfolding lays out the tensor’s values to match how processors read them.
Living Matrix manages the runtime.
Pipeline Parallelism spreads one model across many machines. NukeTorch was born from this pursuit, and it needed control of every layer.
We followed the mathematics, and it delivered what it promised, plus some surprises.
Pipelining is as old as the assembly line, older than the computer. Processors adopted it in the 1960s, and GPipe (2018) made it standard for training large models across many accelerators. Yet it stayed inside large data centers.
At WWDC 2026, Apple fine-tuned Qwen3.5-9B on four M3 Ultra Macs at about 600 tokens/s. NukeTorch ran the same fine-tuning, pipelined across four M5 Max MacBook Pros, at over 5,200 tokens/s. At Apple’s speed, the same 1,000-step fine-tuning would take about 40½ minutes; NukeTorch took under 5 minutes.
Our pipeline parallelism, officially named the Choo-Choo Relay Train, spreads one model across many Macs. Four Macs ran at 93% of four single Macs, and our two-Mac study reached up to 97%. Our next public demo will use 8 to 24 machines.
Pipeline parallelism is here.
NVIDIA’s own robot model, GR00T N1.7, runs faster on four MacBook Pros than on an H100: 2.45× the fastest published H100 result and 4.24× NVIDIA’s own published figure.
We pipelined the 106 GB 4-bit Qwen3.8 Flash Next over four MacBook Pros. NukeTorch measured 177.7 tokens/s for one user and 859.0 tokens/s for 32 users. The closest match we found was 151 tokens/s for one user with Unsloth’s 4-bit model on one NVIDIA H100 NVL. Not a conclusive finding, yet it excites us to keep pursuing NVIDIA on speed.
Next: a series of M5 Ultra machines, now on back-order. At the price of an eight-GPU NVIDIA H100 server, our model projects that Macs will beat NVIDIA in training. It is ‘only’ a projection. But in our experience:
Mathematics does not lie.