Model comparison

Llama3.3-8B-Instruct-Thinking-Heretic-Uncensored-Claude-4.5-Opus-High-Reasoning vs Qwen3-4B

At Q4_K_M, Qwen3-4B is the smaller download — 2,497,280,256 bytes against 4,920,741,056.

From the file· summed bytes, KV per layer

Side by side

Llama3.3-8B-Instruct-Thinking-Heretic-Uncensored-Claude-4.5-Opus-High-ReasoningQwen3-4B
Parameters8.0B4.0B
Architecturellamaqwen3
Layers3236
Native context131,07240,960
Mixture of expertsnono
Quantizations published3729
Smallest quantization1.88 GiB1.01 GiB
Q4_K_M4.58 GiB2.33 GiB
Licenceapache-2.0—

KV cache by context

the term that decides long-context viability
ContextLlama3.3-8B-Instruct-Thinking-Heretic-Uncensored-Claude-4.5-Opus-High-ReasoningQwen3-4BRatio
4,0960.50 GiB0.56 GiB1.13×
8,1921.00 GiB1.13 GiB1.13×
16,3842.00 GiB2.25 GiB1.13×
32,7684.00 GiB4.50 GiB1.13×
65,5368.00 GiB9.00 GiB1.13×
131,07216.00 GiB18.00 GiB1.13×