See what AI models run on 16GB, 32GB, 64GB and 128GB in 2026, with real weight sizes, KV cache costs, quantisation limits and ...