Model comparison
GrammarCoder-7B-Base vs Qwen3-VL-8B-Instruct-abliterated-v1
At Q4_K_M, GrammarCoder-7B-Base is the smaller download — 4,696,210,464 bytes against 5,027,785,792.
From the file· summed bytes, KV per layer
Side by side
| GrammarCoder-7B-Base | Qwen3-VL-8B-Instruct-abliterated-v1 | |
|---|---|---|
| Parameters | 7.6B | 8.8B |
| Architecture | qwen2 | qwen3vl |
| Layers | 28 | — |
| Native context | 131,072 | — |
| Mixture of experts | no | no |
| Quantizations published | 36 | 48 |
| Smallest quantization | 1.78 GiB | 1.97 GiB |
| Q4_K_M | 4.37 GiB | 4.68 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | GrammarCoder-7B-Base | Qwen3-VL-8B-Instruct-abliterated-v1 | Ratio |
|---|---|---|---|
| 4,096 | 0.22 GiB | — | — |
| 8,192 | 0.44 GiB | — | — |
| 16,384 | 0.88 GiB | — | — |
| 32,768 | 1.75 GiB | — | — |
| 65,536 | 3.50 GiB | — | — |
| 131,072 | 7.00 GiB | — | — |