Is granite-3.1-3b-a800m-instruct or HyperCLOVAX-SEED-Text-Instruct-1.5B smaller?

HyperCLOVAX-SEED-Text-Instruct-1.5B is smaller at Q4_K_M: 1.06 GiB against 1.88 GiB.

Which uses less memory at long context?

We cannot resolve the KV layout for one of these models, so we do not publish a comparison.

Model comparison

At Q4_K_M, HyperCLOVAX-SEED-Text-Instruct-1.5B is the smaller download — 1,133,974,368 bytes against 2,016,888,096.

From the file· summed bytes, KV per layer

	granite-3.1-3b-a800m-instruct	HyperCLOVAX-SEED-Text-Instruct-1.5B
Parameters	3.3B	1.6B
Architecture	granitemoe	llama
Layers	32	—
Native context	131,072	—
Mixture of experts	yes, 40 experts	no
Quantizations published	24	1
Smallest quantization	1.16 GiB	1.06 GiB
Q4_K_M	1.88 GiB	1.06 GiB
Licence	apache-2.0	other

the term that decides long-context viability

Context	granite-3.1-3b-a800m-instruct	HyperCLOVAX-SEED-Text-Instruct-1.5B	Ratio
4,096	0.25 GiB	—	—
8,192	0.50 GiB	—	—
16,384	1.00 GiB	—	—
32,768	2.00 GiB	—	—
65,536	4.00 GiB	—	—
131,072	8.00 GiB	—	—