Model comparison

Kimi-Linear-48B-A3B-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 29,702,758,784. At long context the gap widens: Kimi-Linear-48B-A3B-Instruct's KV cache at 32K is 3.2× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

Kimi-Linear-48B-A3B-InstructQwen3-30B-A3B-Thinking-2507
Parameters49.1B30.5B
Architecturekimi-linearqwen3moe
Layers2748
Native context262,144
Mixture of expertsyes, 256 expertsyes, 128 experts
Quantizations published3851
Smallest quantization9.77 GiB7.05 GiB
Q4_K_M27.66 GiB17.28 GiB
Licencemitapache-2.0

KV cache by context

the term that decides long-context viability
ContextKimi-Linear-48B-A3B-InstructQwen3-30B-A3B-Thinking-2507Ratio
4,0960.12 GiB0.38 GiB3.16×
8,1920.24 GiB0.75 GiB3.16×
16,3840.47 GiB1.50 GiB3.16×
32,7680.95 GiB3.00 GiB3.16×
65,5361.90 GiB6.00 GiB3.16×
131,0723.80 GiB12.00 GiB3.16×