The quantization choice makes a huge difference to the phone claim: the 35B repository lists IQ1_M at 8.24 GB and Q4_K_M at 21.2 GB. Those are weight-file sizes, before the KV cache and runtime overhead. MoE reduces how much computation each token needs; the active parameter count isn't the RAM requirement. For a reproducible on-phone example, the exact device, GGUF variant and context length would be more useful than the 35B label alone.