Model comparison
MegaBeam-Mistral-7B-512k vs Qwen3-8B
At Q4_K_M, MegaBeam-Mistral-7B-512k is the smaller download — 4,368,440,224 bytes against 5,027,783,488.
From the file· summed bytes, KV per layer
Side by side
| MegaBeam-Mistral-7B-512k | Qwen3-8B | |
|---|---|---|
| Parameters | 7.2B | 8.2B |
| Architecture | llama | qwen3 |
| Layers | 32 | 36 |
| Native context | 524,288 | 40,960 |
| Mixture of experts | no | no |
| Quantizations published | 19 | 51 |
| Smallest quantization | 2.33 GiB | 2.12 GiB |
| Q4_K_M | 4.07 GiB | 4.68 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | MegaBeam-Mistral-7B-512k | Qwen3-8B | Ratio |
|---|---|---|---|
| 4,096 | 0.50 GiB | 0.56 GiB | 1.13× |
| 8,192 | 1.00 GiB | 1.13 GiB | 1.13× |
| 16,384 | 2.00 GiB | 2.25 GiB | 1.13× |
| 32,768 | 4.00 GiB | 4.50 GiB | 1.13× |
| 65,536 | 8.00 GiB | 9.00 GiB | 1.13× |
| 131,072 | 16.00 GiB | 18.00 GiB | 1.13× |