[concats]

How does it compare for models of any meaningful size?

These 0.6B-4B models are, frankly, just amusing curiosities. But commonly regarded as too error prone for any non-demo work.

The reason why people are buying Apple Silicon today is because the unified memory allows them to run larger models that are cost prohibitive to run otherwise (usually requiring Nvidia server GPUs). It would be much more interesting to see benchmarks for things like Qwen3.5-122B-A10B, GLM-5, or any dense model is the 20b+ range. Thanks.

[LuxBennu]

Agreed. The real value proposition of Apple Silicon for local inference is running models that won't fit on consumer GPUs. I run Qwen 70B 4-bit on an M2 Max 96GB through llama.cpp and it's usable — not fast, but the unified memory means it actually loads. Would be interested to see MetalRT benchmarks at that scale, since the architectural advantages (fused kernels, reduced dispatch overhead) should matter more as models get memory-bandwidth-bound.

[sanchitmonga22]

Fair criticism. Our benchmarks are on small models because MetalRT was built for the voice pipeline use case, where decode latency on 0.6B-4B models is the bottleneck.

You're right that the bigger opportunity on Apple Silicon is large models that don't fit on consumer GPUs. Expanding MetalRT to 7B, 14B, 32B+ is on the roadmap. The architectural advantages(that MetalRT has) should matter even more at that scale where everything becomes memory-bandwidth-bound.

We'll publish benchmarks on larger models as we add support. If you have a specific model/size you'd want to see first, that helps us prioritize.