Model comparison
Neuron-Qwen3-4B-Instruct vs Llama-3.1-8B-Instruct
At Q4_K_M, Neuron-Qwen3-4B-Instruct is the smaller download — 2,497,279,328 bytes against 4,920,739,168.
From the file· summed bytes, KV per layer
Side by side
| Neuron-Qwen3-4B-Instruct | Llama-3.1-8B-Instruct | |
|---|---|---|
| Parameters | 4.0B | 8.0B |
| Architecture | qwen3 | llama |
| Layers | — | 32 |
| Native context | — | 131,072 |
| Mixture of experts | no | no |
| Quantizations published | 36 | 45 |
| Smallest quantization | 0.98 GiB | 2.02 GiB |
| Q4_K_M | 2.33 GiB | 4.58 GiB |
| Licence | — | llama3.1 |
KV cache by context
the term that decides long-context viability
| Context | Neuron-Qwen3-4B-Instruct | Llama-3.1-8B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | — | 0.50 GiB | — |
| 8,192 | — | 1.00 GiB | — |
| 16,384 | — | 2.00 GiB | — |
| 32,768 | — | 4.00 GiB | — |
| 65,536 | — | 8.00 GiB | — |
| 131,072 | — | 16.00 GiB | — |