Not quite true. Depends on number of KV heads. GLM4 32b at IQ4 quant and Q8 context can run full context with only 20GiB VRAM.