Model comparison
reka-flash-3 vs Qwen3-30B-A3B-Thinking-2507
At Q4_K_M, reka-flash-3 is the smaller download — 13,610,362,336 bytes against 18,556,685,824. At long context the gap widens: Qwen3-30B-A3B-Thinking-2507's KV cache at 32K is 1.4× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| reka-flash-3 | Qwen3-30B-A3B-Thinking-2507 | |
|---|---|---|
| Parameters | 20.9B | 30.5B |
| Architecture | llama | qwen3moe |
| Layers | 44 | 48 |
| Native context | 32,768 | 262,144 |
| Mixture of experts | no | yes, 128 experts |
| Quantizations published | 38 | 51 |
| Smallest quantization | 6.88 GiB | 7.05 GiB |
| Q4_K_M | 12.68 GiB | 17.28 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | reka-flash-3 | Qwen3-30B-A3B-Thinking-2507 | Ratio |
|---|---|---|---|
| 4,096 | 0.52 GiB | 0.38 GiB | 1.38× |
| 8,192 | 1.03 GiB | 0.75 GiB | 1.38× |
| 16,384 | 2.06 GiB | 1.50 GiB | 1.38× |
| 32,768 | 4.13 GiB | 3.00 GiB | 1.38× |
| 65,536 | 8.25 GiB | 6.00 GiB | 1.38× |
| 131,072 | 16.50 GiB | 12.00 GiB | 1.38× |