Model comparison
Atomight-V2.5-1.7B vs Llama-3.2-3B-Instruct
At Q4_K_M, Atomight-V2.5-1.7B is the smaller download — 1,107,410,112 bytes against 2,019,377,440.
From the file· summed bytes, KV per layer
Side by side
| Atomight-V2.5-1.7B | Llama-3.2-3B-Instruct | |
|---|---|---|
| Parameters | 1.7B | 3.2B |
| Architecture | qwen3 | llama |
| Layers | 28 | 28 |
| Native context | 40,960 | 131,072 |
| Mixture of experts | no | no |
| Quantizations published | 36 | 39 |
| Smallest quantization | 0.48 GiB | 0.85 GiB |
| Q4_K_M | 1.03 GiB | 1.88 GiB |
| Licence | mit | — |
KV cache by context
the term that decides long-context viability
| Context | Atomight-V2.5-1.7B | Llama-3.2-3B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | 0.44 GiB | 0.44 GiB | 1.00× |
| 8,192 | 0.88 GiB | 0.88 GiB | 1.00× |
| 16,384 | 1.75 GiB | 1.75 GiB | 1.00× |
| 32,768 | 3.50 GiB | 3.50 GiB | 1.00× |
| 65,536 | 7.00 GiB | 7.00 GiB | 1.00× |
| 131,072 | 14.00 GiB | 14.00 GiB | 1.00× |