Hardware · Rockchip
What AI models can the Rockchip RK3588 / RK3588S (4 GB) run?
- RAM
- 4.0 GB
- Usable for models
- 2.8 GB
- Memory bandwidth
- 34 GB/s
- Runs models on
- CPU cores
- Released
- 2022
Models run on the CPU cores here, so the memory’s speed and how fast those cores can read it set the pace: small models (about 1-4B) are usable, bigger ones are slow. We plan around 2.8 GB of the 4.0 GB (what Linux leaves free for a model). 6 TOPS NPU (three cores). llama.cpp and Ollama use the CPU. Rockchip’s RKLLM toolkit (and front ends such as RKLLama) runs converted 8-bit models on the NPU; generation is still limited by memory speed, so it isn’t much faster than the CPU, but it leaves the CPU free.
More about this hardware
| Maker | Rockchip |
|---|---|
| Type | Single-board computer or edge AI board |
| Share of memory for models | 70% (what Linux typically leaves free on a board this size) |
| Chip | Rockchip RK3588 |
| AI accelerator | 6 TOPS NPU (three cores). llama.cpp and Ollama use the CPU. Rockchip’s RKLLM toolkit (and front ends such as RKLLama) runs converted 8-bit models on the NPU; generation is still limited by memory speed, so it isn’t much faster than the CPU, but it leaves the CPU free. |
| Boards with this chip | Orange Pi 5 / 5 Plus / 5 Max / 5 Ultra · Radxa ROCK 5B / 5B+ / 5 ITX · FriendlyElec NanoPC-T6 · Khadas Edge2 · Turing RK1 · Vicharak Axon · KICKPI K8C · Seeed reComputer RK3588 |
| Memory type | LPDDR4X-4224, 64-bit (boards with LPDDR5 reach about 44 GB/s) |
| Bandwidth figure | 33.8 GB/s, the chip’s rated memory speed |
| Note | 4× Cortex-A76 at 2.4 GHz plus 4× Cortex-A55, Mali-G610 GPU. Speeds assume llama.cpp on the CPU: Turing Pi measured 11.25 tokens/s for Llama 3.2 3B at Q4_K_M on an RK3588, about 60% of the memory’s rated speed. The RK3588S is the same chip with fewer I/O lanes. Use a fan or a large heatsink. |
| Specifications | www.orangepi.org |
Apps that run models well on the Rockchip RK3588 / RK3588S (4 GB)
Linux on ARM boards
2 products with the Rockchip RK3588 / RK3588S (4 GB)
Best LLMs for the Rockchip RK3588 / RK3588S (4 GB)
| Quality | Model | Quant | Memory | Speed | Runs |
|---|---|---|---|---|---|
| 53Fair quality (C) · 53 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 74/100 (quality, fit, speed and memory use) sets the ranking. | LFM3 1.7B | Q4_K_M | 2.0 GB | 20.0 tok/s | Runs well |
| 57Good quality (B) · 57 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 73/100 (quality, fit, speed and memory use) sets the ranking. | MiniCPM5 2B | Q4_K_M | 2.7 GB | 13.7 tok/s | Runs welltight fit |
| 47Fair quality (C) · 47 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 71/100 (quality, fit, speed and memory use) sets the ranking. | LFM3-VL 2B | Q4_K_M | 2.1 GB | 18.0 tok/s | Runs well |
| 39Basic quality (D) · 39 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 70/100 (quality, fit, speed and memory use) sets the ranking. | PaddleOCR-VL 1.6 0.9B | Q4_K_M | 1.6 GB | 36.0 tok/s | Runs great |
| 39Basic quality (D) · 39 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 69/100 (quality, fit, speed and memory use) sets the ranking. | MiniCPM-OCR 1.0 1.3B | Q4_K_M | 1.7 GB | 26.6 tok/s | Runs well |
| 46Fair quality (C) · 46 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 68/100 (quality, fit, speed and memory use) sets the ranking. | Spark-X2.6 4B | Q2_K | 2.8 GB | 13.8 tok/s | Runs welltight fit |
| 38Basic quality (D) · 38 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 68/100 (quality, fit, speed and memory use) sets the ranking. | MiniCPM-V 4.7 1B | Q4_K_M | 1.9 GB | 22.3 tok/s | Runs well |
| 30Basic quality (D) · 30 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking. | Unlimited OCR 2 (3B-A0.6B) | Q2_K | 2.7 GB | 73.0 tok/s | Runs welltight fit |
| 38Basic quality (D) · 38 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking. | TeleOCR | Q4_K_M | 2.6 GB | 24.3 tok/s | Runs welltight fit |
| 34Basic quality (D) · 34 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking. | MiniCPM5 1B | Q4_K_M | 1.7 GB | 32.0 tok/s | Runs great |
| 38Basic quality (D) · 38 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking. | TMax 2B | Q4_K_M | 2.1 GB | 18.4 tok/s | Runs well |
| 42Fair quality (C) · 42 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 66/100 (quality, fit, speed and memory use) sets the ranking. | TMax 4B | Q2_K | 2.8 GB | 13.5 tok/s | Runs welltight fit |
Best Image models for the Rockchip RK3588 / RK3588S (4 GB)
| Quality | Model | Quant | Memory | Speed | Runs |
|---|---|---|---|---|---|
| 28Basic quality (D) · 28 of 100: how capable the model is at FP16, whatever it runs on. Match score 59/100 (quality, fit, speed and memory use) sets the ranking. | SD Turbo | FP16 | 2.6 GB | 12 s | Runs welltight fit · encoder unloaded after use |
| 42Fair quality (C) · 42 of 100: how capable the model is at FP16, whatever it runs on. Match score 55/100 (quality, fit, speed and memory use) sets the ranking. | Realistic Vision V5.1 | FP16 | 2.5 GB | 3.1 min | Very slowtight fit · encoder unloaded after use |
| 40Fair quality (C) · 40 of 100: how capable the model is at FP16, whatever it runs on. Match score 54/100 (quality, fit, speed and memory use) sets the ranking. | DreamShaper 8 | FP16 | 2.5 GB | 3.1 min | Very slowtight fit · encoder unloaded after use |
| 30Basic quality (D) · 30 of 100: how capable the model is at FP16, whatever it runs on. Match score 54/100 (quality, fit, speed and memory use) sets the ranking. | LCM DreamShaper v7 | FP16 | 2.5 GB | 38 s | Runs welltight fit · encoder unloaded after use |
| 50Fair quality (C) · 50 of 100: how capable the model is at FP16, whatever it runs on. Match score 52/100 (quality, fit, speed and memory use) sets the ranking. | PixArt-Σ XL 1024 | FP16 | 2.3 GB | 4.8 min | Very slowCPU offload |
| 35Basic quality (D) · 35 of 100: how capable the model is at FP16, whatever it runs on. Match score 52/100 (quality, fit, speed and memory use) sets the ranking. | Stable Diffusion 1.5 | FP16 | 2.5 GB | 3.1 min | Very slowtight fit · encoder unloaded after use |
Best Video models for the Rockchip RK3588 / RK3588S (4 GB)
| Quality | Model | Quant | Memory | Speed | Runs |
|---|---|---|---|---|---|
| 35Basic quality (D) · 35 of 100: how capable the model is at FP16, whatever it runs on. Match score 44/100 (quality, fit, speed and memory use) sets the ranking. | AnimateDiff v3 (SD 1.5) | FP16 | 2.6 GB | 63.5 min | Very slowtight fit · CPU offload |
Best Decision models for the Rockchip RK3588 / RK3588S (4 GB)
| Quality | Model | Quant | Memory | Speed | Runs |
|---|---|---|---|---|---|
| 78Great quality (A) · 78 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 85/100 (quality, fit, speed and memory use) sets the ranking. | OpenAI Privacy Filter | Q8_0 | 2.4 GB | 120 ms | Runs welltight fit |
| 86Excellent quality (A+) · 86 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 81/100 (quality, fit, speed and memory use) sets the ranking. | Jina Reranker v3.5 | Q8_0 | 1.7 GB | 2.6 s | Very slow |
| 84Great quality (A) · 84 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 80/100 (quality, fit, speed and memory use) sets the ranking. | Jina Reranker v3 | Q8_0 | 1.7 GB | 2.6 s | Very slow |
| 80Great quality (A) · 80 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 80/100 (quality, fit, speed and memory use) sets the ranking. | GLiNER2.5 Decide | Q8_0 | 1.2 GB | 1.1 s | Slow |
| 80Great quality (A) · 80 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 79/100 (quality, fit, speed and memory use) sets the ranking. | FRIDA-Decisions | Q8_0 | 1.7 GB | 1.9 s | Very slow |
| 80Great quality (A) · 80 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 78/100 (quality, fit, speed and memory use) sets the ranking. | Qwen3 Reranker 0.6B | Q8_0 | 1.7 GB | 2.6 s | Very slow |
| 68Good quality (B) · 68 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 78/100 (quality, fit, speed and memory use) sets the ranking. | Llama Prompt Guard 2 22M | Q8_0 | 0.7 GB | 333 ms | Runs well |
| 81Great quality (A) · 81 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 78/100 (quality, fit, speed and memory use) sets the ranking. | mxbai-rerank-large v2 | Q6_K | 2.1 GB | 6.6 s | Very slow |
Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.