Model comparison
Teuken-7B-instruct-research-v0.4 vs Qwen3-4B
At Q4_K_M, Qwen3-4B is the smaller download — 2,497,280,256 bytes against 5,018,868,320. At long context the gap widens: Teuken-7B-instruct-research-v0.4's KV cache at 32K is 4.5× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Teuken-7B-instruct-research-v0.4 | Qwen3-4B | |
|---|---|---|
| Parameters | 7.5B | 4.0B |
| Architecture | llama | qwen3 |
| Layers | 32 | 36 |
| Native context | 4,096 | 40,960 |
| Mixture of experts | no | no |
| Quantizations published | 41 | 29 |
| Smallest quantization | 2.48 GiB | 1.01 GiB |
| Q4_K_M | 4.67 GiB | 2.33 GiB |
| Licence | other | — |
KV cache by context
the term that decides long-context viability
| Context | Teuken-7B-instruct-research-v0.4 | Qwen3-4B | Ratio |
|---|---|---|---|
| 4,096 | 0.13 GiB | 0.56 GiB | 4.50× |
| 8,192 | 0.25 GiB | 1.13 GiB | 4.50× |
| 16,384 | 0.50 GiB | 2.25 GiB | 4.50× |
| 32,768 | 1.00 GiB | 4.50 GiB | 4.50× |
| 65,536 | 2.00 GiB | 9.00 GiB | 4.50× |
| 131,072 | 4.00 GiB | 18.00 GiB | 4.50× |