Model comparison
Kimi-VL-A3B-Thinking-2506 vs Qwen3.8-27B
At Q4_K_M, Kimi-VL-A3B-Thinking-2506 is the smaller download — 10,540,747,680 bytes against 16,810,714,336. At long context the gap widens: Kimi-VL-A3B-Thinking-2506's KV cache at 32K is 2.1× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Kimi-VL-A3B-Thinking-2506 | Qwen3.8-27B | |
|---|---|---|
| Parameters | 16.4B | 27.8B |
| Architecture | deepseek2 | qwen35 |
| Layers | 27 | 64 |
| Native context | 131,072 | 262,144 |
| Mixture of experts | yes, 64 experts | no |
| Quantizations published | 21 | 22 |
| Smallest quantization | 6.13 GiB | 8.39 GiB |
| Q4_K_M | 9.82 GiB | 15.66 GiB |
| Licence | — | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Kimi-VL-A3B-Thinking-2506 | Qwen3.8-27B | Ratio |
|---|---|---|---|
| 4,096 | 0.12 GiB | 0.25 GiB | 2.11× |
| 8,192 | 0.24 GiB | 0.50 GiB | 2.11× |
| 16,384 | 0.47 GiB | 1.00 GiB | 2.11× |
| 32,768 | 0.95 GiB | 2.00 GiB | 2.11× |
| 65,536 | 1.90 GiB | 4.00 GiB | 2.11× |
| 131,072 | 3.80 GiB | 8.00 GiB | 2.11× |