Rockchip RK3566

Hardware · Rockchip

What AI models can the Rockchip RK3566 / RK3568 (2 GB) run?

RAM
2.0 GB
Usable for models
1.2 GB
Memory bandwidth
13 GB/s
Runs models on
CPU cores
Released
2021
36 / 622 LLMs run Best: SmolLM2 135M Instruct 0 / 61 Image models run None fit 0 / 43 Video models run None fit 25 / 70 Decision models run Best: Llama Prompt Guard 2 86M

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 1.2 GB of the 2.0 GB (what Linux leaves free for a model). 0.8 TOPS NPU (about 1 TOPS on the RK3568) for small vision models through Rockchip’s RKNN toolkit.

More about this hardware

MakerRockchip
TypeSingle-board computer or edge AI board
Share of memory for models60% (what Linux typically leaves free on a board this size)
ChipRockchip RK3566
AI accelerator0.8 TOPS NPU (about 1 TOPS on the RK3568) for small vision models through Rockchip’s RKNN toolkit.
Boards with this chipRadxa ROCK 3A / 3C / ZERO 3 · Orange Pi 3B · Pine64 Quartz64 · FriendlyElec NanoPi R5S
Memory typeLPDDR4/LPDDR4X or DDR4 up to 3200 MT/s, 32-bit
Bandwidth figure12.8 GB/s, the chip’s rated memory speed
Note4× Cortex-A55 at 1.8 GHz (RK3566) or 2.0 GHz (RK3568), Mali-G52 2EE GPU, up to 8 GB. Entry level: small models only, at a few tokens per second. Four A55 cores can’t read memory at anything near its rated speed, so we assume 20% (estimate).
Specificationswww.cnx-software.com

Apps that run models well on the Rockchip RK3566 / RK3568 (2 GB)

Linux on ARM boards

  1. OllamaOne command with Ollama’s own build of this model.
  2. LM StudioPoint-and-click: search, download and chat in one window.
  3. llama.cppUsually the first to support new model architectures.
  4. Docker Model RunnerHandy if you already work in Docker.

For a typical model that fits in memory. Each model page picks the best app for that model on this hardware. All apps

Best LLMs for the Rockchip RK3566 / RK3568 (2 GB)

Quality Model Quant Memory Speed Runs
3Basic quality (D) · 3 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 56/100 (quality, fit, speed and memory use) sets the ranking. SmolLM2 135M InstructHugging Face Q3_K_M 0.9 GB 43.4 tok/s Runs great
7Basic quality (D) · 7 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 56/100 (quality, fit, speed and memory use) sets the ranking. LFM2.5 230MLiquid AI Q2_K 0.9 GB 31.2 tok/s Runs great
2Basic quality (D) · 2 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 56/100 (quality, fit, speed and memory use) sets the ranking. Falcon-H1 Tiny 90M InstructTII Q5_K_M 0.9 GB 40.1 tok/s Runs great
1Basic quality (D) · 1 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 55/100 (quality, fit, speed and memory use) sets the ranking. SmolLM 135M InstructHugging Face Q3_K_M 0.8 GB 43.4 tok/s Runs great
9Basic quality (D) · 9 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 55/100 (quality, fit, speed and memory use) sets the ranking. LFM2.5 350MLiquid AI Q2_K 0.9 GB 20.5 tok/s Runs well
8Basic quality (D) · 8 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 54/100 (quality, fit, speed and memory use) sets the ranking. Granite 4.0 H 350MIBM Q2_K 0.9 GB 21.1 tok/s Runs well
8Basic quality (D) · 8 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 54/100 (quality, fit, speed and memory use) sets the ranking. LFM2 350MLiquid AI Q2_K 0.9 GB 20.5 tok/s Runs well
4Basic quality (D) · 4 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 54/100 (quality, fit, speed and memory use) sets the ranking. SmolVLM 256M InstructHugging Face Q2_K 1.0 GB 27.6 tok/s Runs well
4Basic quality (D) · 4 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 54/100 (quality, fit, speed and memory use) sets the ranking. Gemma 3 270MGoogle Q2_K 0.8 GB 26.5 tok/s Runs well
18Basic quality (D) · 18 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 53/100 (quality, fit, speed and memory use) sets the ranking. Qwen3.5 0.8BAlibaba Q2_K 1.1 GB 8.2 tok/s Slowtight fit
8Basic quality (D) · 8 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 53/100 (quality, fit, speed and memory use) sets the ranking. LFM2-VL 450MLiquid AI Q2_K 1.0 GB 15.9 tok/s Runs well
4Basic quality (D) · 4 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 52/100 (quality, fit, speed and memory use) sets the ranking. ERNIE 4.5 0.3BBaidu Q2_K 1.0 GB 19.9 tok/s Runs well

Best Image models for the Rockchip RK3566 / RK3568 (2 GB)

None of the image models in our catalog fit in this much memory.

Best Video models for the Rockchip RK3566 / RK3568 (2 GB)

None of the video models in our catalog fit in this much memory.

Best Decision models for the Rockchip RK3566 / RK3568 (2 GB)

Quality Model Quant Memory Speed Runs
76Great quality (A) · 76 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking. Llama Prompt Guard 2 86MMeta Q8_0 1.0 GB 13.0 s Very slow
75Great quality (A) · 75 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 73/100 (quality, fit, speed and memory use) sets the ranking. mxbai-rerank-base v2Mixedbread Q6_K 1.1 GB 21.1 s Very slowtight fit
74Great quality (A) · 74 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 73/100 (quality, fit, speed and memory use) sets the ranking. GLiClass ModernBERT Large v3Knowledgator Q8_0 1.1 GB 9.3 s Very slowtight fit
70Great quality (A) · 70 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 73/100 (quality, fit, speed and memory use) sets the ranking. GTE Reranker ModernBERT BaseAlibaba Q8_0 0.8 GB 7.0 s Very slow
68Good quality (B) · 68 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 73/100 (quality, fit, speed and memory use) sets the ranking. Llama Prompt Guard 2 22MMeta Q8_0 0.7 GB 3.3 s Very slow
72Great quality (A) · 72 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 72/100 (quality, fit, speed and memory use) sets the ranking. DeBERTa v3 Large Zero-shot v2.0Moritz Laurer Q8_0 1.2 GB 10.2 s Very slowtight fit
70Great quality (A) · 70 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 72/100 (quality, fit, speed and memory use) sets the ranking. Jina Reranker v2 Base MultilingualJina AI Q8_0 1.0 GB 13.0 s Very slow
68Good quality (B) · 68 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 72/100 (quality, fit, speed and memory use) sets the ranking. Granite Embedding Reranker English R2IBM Q8_0 0.8 GB 7.0 s Very slow

Best LLMs by task on this hardware Open in Model Finder

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.

Was this page helpful?