Model comparison
Moonlight-16B-A3B-Instruct vs Qwen3-Coder-30B-A3B-Instruct
At Q4_K_M, Moonlight-16B-A3B-Instruct is the smaller download — 10,537,205,632 bytes against 18,556,689,568. At long context the gap widens: Moonlight-16B-A3B-Instruct's KV cache at 32K is 3.2× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Moonlight-16B-A3B-Instruct | Qwen3-Coder-30B-A3B-Instruct | |
|---|---|---|
| Parameters | 16.0B | 30.5B |
| Architecture | deepseek2 | qwen3moe |
| Layers | 27 | 48 |
| Native context | 8,192 | 262,144 |
| Mixture of experts | yes, 64 experts | yes, 128 experts |
| Quantizations published | 48 | 46 |
| Smallest quantization | 4.77 GiB | 7.46 GiB |
| Q4_K_M | 9.81 GiB | 17.28 GiB |
| Licence | mit | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Moonlight-16B-A3B-Instruct | Qwen3-Coder-30B-A3B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | 0.12 GiB | 0.38 GiB | 3.16× |
| 8,192 | 0.24 GiB | 0.75 GiB | 3.16× |
| 16,384 | 0.47 GiB | 1.50 GiB | 3.16× |
| 32,768 | 0.95 GiB | 3.00 GiB | 3.16× |
| 65,536 | 1.90 GiB | 6.00 GiB | 3.16× |
| 131,072 | 3.80 GiB | 12.00 GiB | 3.16× |