Mistral AI's open-weight flagship, released in October: a 1T mixture-of-experts model with about 52B active per token. It reads text and images (the vision encoder is a separate 3.4 GB file) and has a 1M-token window. Its compressed attention cache costs about 82 KB per token, so context is cheap next to the weights. Only the 2-bit files fit any machine in this index, a 512 GB Mac Studio: IQ2_XXS (ggml-org's 298 GB file) needs about 311 GB at 4K and 334 GB at 256K, both comfortable, and about 401 GB across the full 1M window, which is tight. Q2_K (381 GB) needs about 398 GB at 4K and is tight on that Mac up to 64K. Q4_K_M (635 GB, about 664 GB at 4K) and Q8_0 (1.1 TB) need multi-GPU servers. llama.cpp runs it since late October. Sizes come from ggml-org's GGUF files, and no per-level quality loss has been published.
Mistral’s 1‑Trillion‑Parameter Model Is Shattering AI Cost & Performance—See Why It’s a Game‑Changer