Boards / Rockchip RK3576
RK3576 boards (NPU): LLM benchmarks
Rockchip's mid-range SoC with an NPU, found on boards like the Radxa ROCK 4D.
At a glance
- Measured runs58
- Models tested20
- Fastest run51.22 tok/s
- LFM2.5-VL 450M, w4a16, RKLLM v1.3.1 · source ↗
- Sources1
Where to buy
- Radxa ROCK 4D 8GB: $106 listed price, source ↗ (2026-10-17) Check price on Amazon
Every run on this board
| Model | Quant | Runtime | Gen tok/s | Prompt tok/s | TTFT ms | Source |
|---|---|---|---|---|---|---|
| LFM2.5-VL 450M | w4a16 | RKLLM v1.3.1 | 51.22 | – | 246.03 | source ↗ commit 2026-09-29 |
| LFM2.5-VL 450M | w4a16_g128 | RKLLM v1.3.1 | 48.65 | – | 260.98 | source ↗ commit 2026-09-29 |
| MiniCPM4 0.5B | w4a16 | RKLLM v1.3.1 | 34.57 | – | 330.8 | source ↗ commit 2026-09-29 |
| LFM2.5-VL 450M | w8a8 | RKLLM v1.3.1 | 32.75 | – | 239.34 | source ↗ commit 2026-09-29 |
| MiniCPM4 0.5B | w4a16_g128 | RKLLM v1.3.1 | 32.66 | – | 359.16 | source ↗ commit 2026-09-29 |
| Qwen2 0.5B | w4a16 | RKLLM v1.3.1 | 32.56 | – | 342.37 | source ↗ commit 2026-09-29 |
| Qwen2 0.5B | w4a16_g128 | RKLLM v1.3.1 | 30.33 | – | 361.78 | source ↗ commit 2026-09-29 |
| Qwen3 0.6B | w4a16 | RKLLM v1.3.1 | 24.85 | – | 468.61 | source ↗ commit 2026-09-29 |
| DeepSeek-OCR 3B(A570M) | w4a16 | RKLLM v1.3.1 | 24.38 | – | 1034.97 | source ↗ commit 2026-09-29 |
| Qwen3 0.6B | w4a16_g128 | RKLLM v1.3.1 | 23.48 | – | 506.41 | source ↗ commit 2026-09-29 |
| MiniCPM4 0.5B | w8a8 | RKLLM v1.3.1 | 23.28 | – | 321.12 | source ↗ commit 2026-09-29 |
| Qwen2 0.5B | w8a8 | RKLLM v1.3.1 | 21.73 | – | 344.27 | source ↗ commit 2026-09-29 |
| TinyLlama 1.1B | w4a16 | RKLLM v1.3.1 | 19.71 | – | 543.68 | source ↗ commit 2026-09-29 |
| Qwen3.5 0.8B | w4a16 | RKLLM v1.3.1 | 18.79 | – | 1369.31 | source ↗ commit 2026-09-29 |
| TinyLlama 1.1B | w4a16_g128 | RKLLM v1.3.1 | 18.08 | – | 672.61 | source ↗ commit 2026-09-29 |
| Qwen3.5 0.8B | w4a16_g128 | RKLLM v1.3.1 | 17.88 | – | 1392.09 | source ↗ commit 2026-09-29 |
| DeepSeek-OCR 3B(A570M) | w8a8 | RKLLM v1.3.1 | 17.26 | – | 1242.78 | source ↗ commit 2026-09-29 |
| Qwen3 0.6B | w8a8 | RKLLM v1.3.1 | 17.17 | – | 461.54 | source ↗ commit 2026-09-29 |
| Spark-X2.5 1.7B | w4a16 | RKLLM v1.3.1 | 13.57 | – | 768.66 | source ↗ commit 2026-09-29 |
| InternLM2 1.8B | w4a16 | RKLLM v1.3.1 | 13.55 | – | 765.92 | source ↗ commit 2026-09-29 |
| Qwen3.5 0.8B | w8a8 | RKLLM v1.3.1 | 13.18 | – | 1342.54 | source ↗ commit 2026-09-29 |
| Qwen3-VL 2B | w4a16 | RKLLM v1.3.1 | 12.68 | – | 802.94 | source ↗ commit 2026-09-29 |
| TinyLlama 1.1B | w8a8 | RKLLM v1.3.1 | 12.18 | – | 534.13 | source ↗ commit 2026-09-29 |
| Spark-X2.5 1.7B | w4a16_g128 | RKLLM v1.3.1 | 12.14 | – | 981.5 | source ↗ commit 2026-09-29 |
| InternLM2 1.8B | w4a16_g128 | RKLLM v1.3.1 | 12.05 | – | 986.14 | source ↗ commit 2026-09-29 |
| Qwen3-VL 2B | w4a16_g128 | RKLLM v1.3.1 | 11.42 | – | 983.2 | source ↗ commit 2026-09-29 |
| Qwen3.5 2B | w4a16 | RKLLM v1.3.1 | 11.03 | – | 1661.01 | source ↗ commit 2026-09-29 |
| RWKV7 1.5B | w4a16 | RKLLM v1.3.1 | 10.18 | – | 1572.09 | source ↗ commit 2026-09-29 |
| Qwen3.5 2B | w4a16_g128 | RKLLM v1.3.1 | 10.1 | – | 1852.42 | source ↗ commit 2026-09-29 |
| RWKV7 1.5B | w4a16_g128 | RKLLM v1.3.1 | 9.49 | – | 1733.08 | source ↗ commit 2026-09-29 |
| Gemma 4 E2B | w4a16 | RKLLM v1.3.1 | 9.23 | – | 1219.25 | source ↗ commit 2026-09-29 |
| Gemma 2 2B | w4a16 | RKLLM v1.3.1 | 9.14 | – | 997.92 | source ↗ commit 2026-09-29 |
| TeleChat2 3B | w4a16 | RKLLM v1.3.1 | 8.97 | – | 1153.96 | source ↗ commit 2026-09-29 |
| Gemma 2 2B | w4a16_g128 | RKLLM v1.3.1 | 8.27 | – | 1312.15 | source ↗ commit 2026-09-29 |
| Gemma 4 E2B | w4a16_g128 | RKLLM v1.3.1 | 8.27 | – | 1445.36 | source ↗ commit 2026-09-29 |
| TeleChat2 3B | w4a16_g128 | RKLLM v1.3.1 | 8.14 | – | 1408.37 | source ↗ commit 2026-09-29 |
| Spark-X2.5 1.7B | w8a8 | RKLLM v1.3.1 | 7.89 | – | 803.94 | source ↗ commit 2026-09-29 |
| InternLM2 1.8B | w8a8 | RKLLM v1.3.1 | 7.86 | – | 750.31 | source ↗ commit 2026-09-29 |
| Gemma 3n 2B | w4a16 | RKLLM v1.3.1 | 7.78 | – | 3260.05 | source ↗ commit 2026-09-29 |
| Qwen3-VL 2B | w8a8 | RKLLM v1.3.1 | 7.62 | – | 785.79 | source ↗ commit 2026-09-29 |
| RWKV7 1.5B | w8a8 | RKLLM v1.3.1 | 7.05 | – | 1534.47 | source ↗ commit 2026-09-29 |
| Qwen3.5 2B | w8a8 | RKLLM v1.3.1 | 6.73 | – | 1678.11 | source ↗ commit 2026-09-29 |
| Phi-3 3.8B | w4a16 | RKLLM v1.3.1 | 6.58 | – | 1829.12 | source ↗ commit 2026-09-29 |
| Phi-3 3.8B | w4a16_g128 | RKLLM v1.3.1 | 6.06 | – | 2253.14 | source ↗ commit 2026-09-29 |
| Gemma 4 E2B | w8a8 | RKLLM v1.3.1 | 5.56 | – | 1166.94 | source ↗ commit 2026-09-29 |
| Gemma 2 2B | w8a8 | RKLLM v1.3.1 | 5.24 | – | 956.77 | source ↗ commit 2026-09-29 |
| TeleChat2 3B | w8a8 | RKLLM v1.3.1 | 5.12 | – | 1029.04 | source ↗ commit 2026-09-29 |
| Qwen3.5 4B | w4a16 | RKLLM v1.3.1 | 5.02 | – | 3976.41 | source ↗ commit 2026-09-29 |
| Gemma 3n 2B | w8a8 | RKLLM v1.3.1 | 4.9 | – | 2995.26 | source ↗ commit 2026-09-29 |
| MiniCPM3 4B | w4a16 | RKLLM v1.3.1 | 4.87 | – | 2536.68 | source ↗ commit 2026-09-29 |
| ChatGLM3 6B | w4a16 | RKLLM v1.3.1 | 4.63 | – | 2166.69 | source ↗ commit 2026-09-29 |
| Qwen3.5 4B | w4a16_g128 | RKLLM v1.3.1 | 4.6 | – | 4476.37 | source ↗ commit 2026-09-29 |
| MiniCPM3 4B | w4a16_g128 | RKLLM v1.3.1 | 4.46 | – | 3117.53 | source ↗ commit 2026-09-29 |
| ChatGLM3 6B | w4a16_g128 | RKLLM v1.3.1 | 4.14 | – | 2966.85 | source ↗ commit 2026-09-29 |
| Phi-3 3.8B | w8a8 | RKLLM v1.3.1 | 3.74 | – | 1615.97 | source ↗ commit 2026-09-29 |
| Qwen3.5 4B | w8a8 | RKLLM v1.3.1 | 3.03 | – | 3906.19 | source ↗ commit 2026-09-29 |
| MiniCPM3 4B | w8a8 | RKLLM v1.3.1 | 3.01 | – | 2539.92 | source ↗ commit 2026-09-29 |
| ChatGLM3 6B | w8a8 | RKLLM v1.3.1 | 2.48 | – | 1948.12 | source ↗ commit 2026-09-29 |
Can the Rockchip RK3576 run…
Can it run ChatGLM3 6B?Can it run DeepSeek-OCR 3B?Can it run Gemma 2 2B?Can it run Gemma 3n 2B?Can it run Gemma 4 E2B?Can it run InternLM2 1.8B?Can it run LFM2.5-VL 450M?Can it run MiniCPM3 4B?Can it run MiniCPM4 0.5B?Can it run Phi-3 3.8B?Can it run Qwen2 0.5B?Can it run Qwen3 0.6B?Can it run Qwen3-VL 2B?Can it run Qwen3.5 0.8B?Can it run Qwen3.5 2B?Can it run Qwen3.5 4B?Can it run RWKV7 1.5B?Can it run Spark-X2.5 1.7B?Can it run TeleChat2 3B?Can it run TinyLlama 1.1B?
Have this board? Submit your own run →