Pipeline Parallelism is here: 8x faster than Apple’s WWDC26 fine-tuning

Today we also release NukeTorch for one Mac as a free download, so you can test our speed yourself.

At WWDC 2026, Apple fine-tuned Qwen3.5-9B with MLX: about 180 tokens/s on one M3 Ultra Mac Studio, and about 600 on four. We ran Apple’s recipe, with Apple’s own commands, on MacBook Pros.

The machines differ, and Apple’s figures are the ones spoken in its session. We did not run Apple’s setup. One machine trains 4 examples per step and four machines train 16, as in Apple’s two commands. Every four-Mac run of ours was more than 8× Apple’s figure. The slowest was 8.45×.

On a single 48 GB MacBook Pro, we beat four M3 Ultra machines. We believe we can reasonably say we built the fastest training engine on Apple silicon. For inference the same may be true, but that side is more contested. Our goal is bigger: to beat NVIDIA on speed in training and in inference.

If you know a faster training engine on Apple silicon, show us. We want to see it.

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.

Pipeline parallelism does not make a single prompt faster. Our speed comes from the mathematics. NukeTorch spreads one model across many machines without the usual inefficiency and without special interconnects. Each machine holds its assigned layers of the model, and together they work as one distributed machine.

• It keeps the speed. Four Macs ran at 93% of four single Macs, and no slower than our own data parallel on the same four Macs.
• It needs a fraction of the memory. Qwen3.8 Flash Next is 106.3 GB. Pipelined over four 48 GB MacBook Pros, no Mac used more than 26.6 GB, even with 32 users at once.
• It uses less energy. 7.3% less per trained token than our own data parallelism on the same four Macs. Boring? Please wait until we compare the same model on NVIDIA and AMD.

Our four MacBook Pros were connected by a single Thunderbolt cable between machines. No special interconnect. No rack, no external fans, no special cooling.

Today we used Apple hardware. We will release additional hardware support.

Qwen3.8 Flash Next is 106.3 GB. No single 48 GB Mac can hold it. The pipeline splits it across four Macs, nothing is stored twice, and all four serve every user together.

For reference, the closest match we found is 151 tokens/s for one user with Unsloth’s 4-bit model on one NVIDIA H100 NVL (published). It used different prompts, so it is a reference, not a conclusive finding.

Serving Flash Next to one user, the four whole Macs drew 341.7 W together.

Qwen3.6 35B-A3B, Qwen3.8 27B and NVIDIA’s GR00T N1.7 each fit in one Mac. Here each of the four MacBook Pros runs its own server and serves its own users.

Combined is the sum of every user’s speed. You can watch the screen captures of these runs on four Macs in our Benchmark section.

NVIDIA’s own robot model, GR00T N1.7 at FP16, served 192 clients at once, 48 on each Mac, at 152.07 policy chunks/s. That is 2.45× faster than the fastest published H100 result, and 4.24× faster than NVIDIA’s own published H100 figure. The H100 figures are published by others, and they measure one request at a time.

NukeTorch N1.0.0 is an inference server for one Mac with an OpenAI-compatible API. It is free for individuals, for schools and non-profits, for non-commercial open-source projects, and for businesses with less than US$10 million in revenue and fewer than 250 people.

* Both Qwen models run with a small draft model.

* We used short, easy prompts for these records. Your everyday prompts will more likely run lower. We measured on an M5 Max with 48 GB, and we recommend that as the minimum.

When someone says they are 80% faster than Apple’s own run, it may be hard to believe. We are more than 8×. We need independent reviewers.

If you have the machines, the knowledge and the reputation to back your review, we want to work with you. You need at least four M5 Max or M5 Ultra Macs. If your company needs a local AI data center, even better.

We are a small team, so we have to be selective. Request an evaluation license: [email protected]

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.

Our next public demo will use 8 to 24 machines. Want to see it? Join us.

• Every number here is our own. We seek independent reviewers.
• Apple’s WWDC26 machines differ from ours, and Apple’s figure is a spoken one.
• H100 figures are published by others for one request at a time.
• Beating NVIDIA in training is a projection, not a measurement.
• Engine ships as a binary. Patents pending.

We worked the past 12 months on about four hours of sleep on average. We wanted to bring pipeline parallelism. The speed, efficient memory usage, and less energy were projected and secured.

Today we proved a little. But we believe this technology will change AI data centers worldwide.

Pipeline parallelism is here.

For about ten years I led a team that built a wash sale engine. Wash sale is one of the heaviest number-crunching jobs in Wall Street tax analysis.

Then I met the Transformer. It was a shock. A few billion calculations in a few seconds. To me it was not coding but pure mathematics. I was a frog in a well.

First I tried pipeline parallelism in PyTorch. Nothing in the mathematics blocked it, yet I could not make it work. And I could not trace where it went wrong. So around Halloween last year I started my own Tensor Engine, based on four basic GPU operations. Oh, it was a fun and brutal ride.

We are driven by theories, architectures and benchmarks. We start with the mathematics, then the papers, then the code. Two of those papers are on our Publication page.

For a small team like ours, this would not have been possible without the explosion of frontier AI models. Now we hope to contribute back to the fundamental acceleration of Intelligence.