Is ERNIE-4.5-21B-A3B-Thinking or Qwen3-Coder-30B-A3B-Instruct smaller?

ERNIE-4.5-21B-A3B-Thinking is smaller at Q4_K_M: 12.42 GiB against 17.28 GiB.

Which uses less memory at long context?

ERNIE-4.5-21B-A3B-Thinking — its KV cache at 32K context is 1.7x smaller, because of how its attention layers cache.

Model comparison

ERNIE-4.5-21B-A3B-Thinking vs Qwen3-Coder-30B-A3B-Instruct

At Q4_K_M, ERNIE-4.5-21B-A3B-Thinking is the smaller download — 13,331,018,624 bytes against 18,556,689,568. At long context the gap widens: ERNIE-4.5-21B-A3B-Thinking's KV cache at 32K is 1.7× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	ERNIE-4.5-21B-A3B-Thinking	Qwen3-Coder-30B-A3B-Instruct
Parameters	21.8B	30.5B
Architecture	ernie4_5-moe	qwen3moe
Layers	28	48
Native context	131,072	262,144
Mixture of experts	no	yes, 128 experts
Quantizations published	70	46
Smallest quantization	4.17 GiB	7.46 GiB
Q4_K_M	12.42 GiB	17.28 GiB
Licence	apache-2.0	apache-2.0

KV cache by context

the term that decides long-context viability

Context	ERNIE-4.5-21B-A3B-Thinking	Qwen3-Coder-30B-A3B-Instruct	Ratio
4,096	0.22 GiB	0.38 GiB	1.71×
8,192	0.44 GiB	0.75 GiB	1.71×
16,384	0.88 GiB	1.50 GiB	1.71×
32,768	1.75 GiB	3.00 GiB	1.71×
65,536	3.50 GiB	6.00 GiB	1.71×
131,072	7.00 GiB	12.00 GiB	1.71×