Can I Run Gemma 3 on Beelink SER9 (Ryzen AI 9, 32 GB)?

Written by Jakub Rusinowski · Last updated March 12, 2025

Yes, but it is tight

Yes, but it is tight — Gemma 3 27B Instruct at Q6_K needs about 31.3 GB of the 32 GB usable on Beelink SER9 (Ryzen AI 9, 32 GB), leaving only ~0.7 GB before the runtime starts swapping. Expect ~7.2 tok/s (estimated), with room for about 8,192 tokens of context.

Confidence: medium · Recommended quantization: Q6_K · Estimated speed: ~7.2 tok/s

Beelink SER9 (Ryzen AI 9, 32 GB) — what it gives a model

Usable memory for models	32 GB
Memory bandwidth	256 GB/s
Form factor	Mini PC
Operating system	Windows or Linux
Memory upgradeable	Yes
Price	$859 (lib/data/ai-stations.ts (street price), checked 2026-07-06)

Gemma 3 on Beelink SER9 (Ryzen AI 9, 32 GB): memory by quantization

Quant	Memory needed	Fits 32 GB?	Max context	Est. speed	Download
F16	63.1 GB	✗ No	—	—	54 GB
Q8_0	37.8 GB	✗ No	—	—	28.7 GB
Q6_K	31.3 GB	✓ Yes	8K	~7.2 tok/s	22.1 GB
Q5_K_M	28.3 GB	✓ Yes	8K	~8.1 tok/s	19.1 GB
Q4_K_M	25.4 GB	✓ Yes	8K	~9.2 tok/s	16.3 GB
Q3_K_M	20.6 GB	✓ Yes	16K	~11.9 tok/s	11.5 GB
Q2_K	18 GB	✓ Yes	16K	~14.3 tok/s	8.9 GB

Which Gemma 3 sizes fit

Variant	Needs at Q4_K_M	Fits?	Est. speed
Gemma 3 27B Instruct	25.4 GB	✓ Fits	~9.2 tok/s
Gemma 3 12B Instruct	11.1 GB	✓ Fits	~20.8 tok/s
Gemma 3 4B Instruct	4.4 GB	✓ Fits	~56.2 tok/s
Gemma 3 1B Instruct	2 GB	✓ Fits	~143.6 tok/s

What to watch out for

Only ~0.7 GB of headroom at Q6_K: a longer context or a second application can push this into swapping.

Beelink SER9 32 GB limitations

Shares system memory with the iGPU, so the usable model budget is well below the nominal 32 GB.
Memory bandwidth, not capacity, is the limit here — expect single-digit tokens/sec on large models.

Recommended setup

Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving

How these numbers are calculated

Memory need = quantized weights + KV cache at 8,192 tokens (f16) + 0.8 GB runtime overhead.
32 GB unified memory at 256 GB/s, shared between CPU and GPU.
Throughput is a memory-bandwidth roofline estimate, not a measurement. It is labelled "estimated" everywhere it appears.
KV cache is computed from this model's published attention configuration.

FAQ