Model comparison

Phi-3.5-MoE-instruct vs Qwen3-30B-A3B-Thinking-2507

At Q4_K_M, Qwen3-30B-A3B-Thinking-2507 is the smaller download — 18,556,685,824 bytes against 25,345,994,304.

From the file· summed bytes, KV per layer

Side by side

Phi-3.5-MoE-instructQwen3-30B-A3B-Thinking-2507
Parameters41.9B30.5B
Architecturephimoeqwen3moe
Layers3248
Native context131,072262,144
Mixture of expertsyes, 16 expertsyes, 128 experts
Quantizations published2951
Smallest quantization10.27 GiB7.05 GiB
Q4_K_M23.61 GiB17.28 GiB
Licencemitapache-2.0

KV cache by context

the term that decides long-context viability
ContextPhi-3.5-MoE-instructQwen3-30B-A3B-Thinking-2507Ratio
4,0960.38 GiB
8,1920.75 GiB
16,3841.50 GiB
32,7683.00 GiB
65,5366.00 GiB
131,07212.00 GiB