Have you tried qwen 3.8 flash next? So far I’ve tested it only briefly but it’s definitely a lot better compared to other local llms that I can run (~30B class models). For reference I am on 16 vram and 32 ram, older ddr4 platform.

These “benchmarks” are great for marketing, so not pointless. Just pointless for users.

Afaik they use the same type of memory as phones do, variants of LPDDRx. “Unified” does not denote a different type, it means that both the GPU and the CPU can use the same memory (similarly to integrated GPUs using the main system ram).