heliskyr2

Two things: drop the custom Metal engines — mlx-lm (Apple's MLX) and llama.cpp's Metal backend are already close to the hardware ceiling, and 150 t/s on a hand-rolled engine is genuinely respectable, not a failure. The real limiter on Apple Silicon is memory bandwidth, not compute, so a smaller model at a higher quant frequently gives more t/s than a bigger one at low quant. With 128GB you can comfortably run Qwen2.5-Coder-32B (great cost/quality for coding) or a 4-bit ~70B if you want the ceiling. I keep a small index of Ollama models by size/quant to size-match RAM (linked from p32929.github.io) — handy when deciding what actually fits your bandwidth budget.

Hi, I've been learning Rust for a few days now & I couldn't help but notice the insane built time. then I came to know that Rust was written initially in OCaml and Golang was written in C/C++ & Golang's compile time is just mindblowing. So, I was wondering, if rust was written in C/C++ would the compiler have been faster? Thanks