You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Private chat and code on your Mac. Run a 177B model on a 128 GB M5 Max: local Qwen3.8 Flash-Next inference with a 65,536-token default or optional 131,072-token context, Tokspur's own Metal engine with native MTP or DFlash2 drafting, and an OpenAI-compatible API. Apache 2.0.
Run Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017 locally with Argodrive DS4: NVIDIA CUDA or Apple silicon Metal, NVMe expert streaming, caches and weight preparation. Setup and measured RTX PRO 6000 / M5 Max results included. Based on antirez/ds4; no separate Argodrive app required.
Every AI compute lane in Apple silicon, measured: ANE, GPU tensor units, shader path — numbers, energy, failure modes, and the instruments to re-check them.
LLM inference on a Mac Studio M5 Max (40-core GPU, 48 GB): achieved memory bandwidth (≈550 GB/s, ~90% of 614 GB/s theoretical), decode vs bandwidth bound, mac-llm-bench results. Independent researcher.