Campground makes strong AI run well on the machines people already own; stoke speeds up MLX models on Apple Silicon, checked against the original.
A 1984 Macintosh redrawn, line by line, into a MacBook.
1984 A computer fits on a desk.
2026 A 30-billion-parameter model, at 3-bit, fits on a base M4 with 24 GB.
The code between a model and the GPU leaves speed unused, and every new model, chip and runtime reopens that gap. Narrowing it takes GPU kernels written for each model on each chip: too many to write by hand.
An AI can now draft them. It can’t be trusted to grade them. We build the part that decides what ships.
Given an MLX model, stoke rewrites parts of it as custom Metal GPU kernels, so the model reads short and medium prompts and generates images faster. A kernel ships only if the outputs still match the original and the whole model gets faster.
Qwen3-30B-A3B, faster to first token on a 128-token prompt, through MLX-LM’s own generation call. Apple M4, 24 GB.
Today we run stoke with you, on your model, on the chip you serve on. You get a verified speedup you apply in one line, or a clear answer that nothing beat your baseline.
Campground is built by Shivam Garg.