LLM Roofline
AMD / Beelink / GMKtec · UMA · Backend: llama.cpp

Ryzen AI Max+ 395 Strix Halo (64 GB) — LLM Sizing & Speed Matrix

Strong x86 ROCm/Vulkan alternative to Mac Mini M4 Pro at lower $/GB for 30B-A3B and 32B models.

Reference Pricing
$1,499 Street
Usable Memory
48 GiB (54 tuned)
Memory Bandwidth
256 GB/s (74% eff.)
FP16 Compute
59.4 TFLOPS
Roofline Ridge
156.8 FLOP/B
Power (Load / Idle)
130W / 13W

Buy or Rent Ryzen AI Max+ 395 Strix Halo (64 GB)

Vendor & Cloud Links
GMKtec EVO-X2 (Strix Halo 128 GB)
GMKtec · Direct Hardware
↗
Beelink Strix Halo AI Mini-PC
Beelink · Direct Hardware
↗

All 32 Open-Weight Models on Ryzen AI Max+ 395 Strix Halo (64 GB) (Q4_K_M @ 32k)

Model Status Weights Total @ 32k Decode TTFT (4k) Max Safe Ctx
Llama-3.3 Nemotron Super 49B COMFORTABLE 28.7 GiB 36.5 GiB 5.1 tok/s 14.25s 123k
Mixtral 8x7B Instruct COMFORTABLE 26.9 GiB 32.0 GiB 13.6 tok/s 3.74s 32k
Qwen3 32B COMFORTABLE 18.9 GiB 28.1 GiB 6.6 tok/s 9.67s 40k
Qwen2.5 Coder 32B COMFORTABLE 18.9 GiB 28.1 GiB 6.6 tok/s 9.45s 32k
DeepSeek-R1 Distill Qwen 32B COMFORTABLE 18.9 GiB 28.1 GiB 6.6 tok/s 9.45s 128k
QwQ 32B Reasoning COMFORTABLE 18.9 GiB 28.1 GiB 6.6 tok/s 9.45s 40k
GLM-4 32B 0414 COMFORTABLE 18.6 GiB 21.9 GiB 8.6 tok/s 9.37s 32k
Qwen3 30B-A3B COMFORTABLE 17.6 GiB 21.4 GiB 31.7 tok/s 1.17s 40k
Gemma 3 27B IT COMFORTABLE 15.8 GiB 20.0 GiB 9.4 tok/s 7.88s 128k
Mistral Small 3.1 24B COMFORTABLE 13.8 GiB 20.1 GiB 9.4 tok/s 6.84s 128k
Codestral 22B COMFORTABLE 12.8 GiB 21.2 GiB 8.9 tok/s 6.55s 32k
Qwen3 14B COMFORTABLE 8.5 GiB 14.8 GiB 13.1 tok/s 4.35s 40k
Qwen2.5 14B Instruct COMFORTABLE 8.5 GiB 15.8 GiB 12.2 tok/s 4.39s 32k
DeepSeek-R1 Distill Qwen 14B COMFORTABLE 8.5 GiB 15.8 GiB 12.2 tok/s 4.39s 128k
Phi-4 14B COMFORTABLE 8.5 GiB 12.8 GiB 15.2 tok/s 4.32s 16k
Gemma 3 12B IT COMFORTABLE 7.0 GiB 10.4 GiB 18.9 tok/s 3.62s 128k
Qwen3 8B COMFORTABLE 4.7 GiB 10.3 GiB 19.1 tok/s 2.46s 40k
Llama 3.1 8B Instruct COMFORTABLE 4.6 GiB 9.7 GiB 20.5 tok/s 2.40s 128k
Qwen2.5 7B Instruct COMFORTABLE 4.4 GiB 7.2 GiB 28.8 tok/s 2.25s 32k
Gemma 3 4B IT COMFORTABLE 2.5 GiB 4.2 GiB 52.9 tok/s 1.30s 128k
Qwen2.5 72B Instruct TUNABLE_UMA 41.9 GiB 53.5 GiB 3.4 tok/s 20.83s 32k
DeepSeek-R1 Distill Llama 70B TUNABLE_UMA 40.7 GiB 52.3 GiB 3.5 tok/s 20.25s 37k
Llama 3.3 70B Instruct TUNABLE_UMA 40.7 GiB 52.3 GiB 3.5 tok/s 20.25s 37k
Llama 3.1 70B Instruct TUNABLE_UMA 40.7 GiB 52.3 GiB 3.5 tok/s 20.25s 37k
DeepSeek-R1 671B OOM 386.4 GiB 390.1 GiB OOM — OOM
DeepSeek-V3 671B OOM 386.4 GiB 390.1 GiB OOM — OOM
Llama 3.1 405B Instruct OOM 233.8 GiB 252.3 GiB OOM — OOM
Llama 4 Maverick 400B-A17B OOM 231.5 GiB 238.8 GiB OOM — OOM
Qwen3 235B-A22B OOM 135.4 GiB 142.4 GiB OOM — OOM
Mixtral 8x22B Instruct OOM 81.2 GiB 89.6 GiB OOM — OOM
Command A 111B OOM 63.9 GiB 68.9 GiB OOM — OOM
Llama 4 Scout 109B-A17B OOM 62.8 GiB 70.0 GiB OOM — OOM

Interactive Ryzen AI Max+ 395 Strix Halo (64 GB) Simulator

49.9B Dense · GQA
256 GB/s · 59.4 TFLOPS
3. Weight Quantization 4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 92k (123k tuned)
Memory Allocation Footprint 36.54 / 48.00 GiB usable
Weights: 28.74 GiB KV Cache: 6.13 GiB Overhead: 1.68 GiB Headroom: 11.46 GiB
Decode Speed
5.1
tok/s (batch=1)
Prefill Speed
288
prompt tok/s
TTFT Latency
14.25s
for 4,096 tok
Active Params
49.9B
of 49.9B total
llama.cpp Server Command 
llama-server -hf nvidia/Llama-3_3-Nemotron-Super-49B-v1:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080
Context KV Cache Total VRAM Decode Status
4k 0.77 GiB 30.62 GiB 6.0 tok/s COMFORTABLE
8k 1.53 GiB 31.94 GiB 5.8 tok/s COMFORTABLE
16k 3.06 GiB 33.48 GiB 5.5 tok/s COMFORTABLE
32k 6.13 GiB 36.54 GiB 5.1 tok/s COMFORTABLE
64k 12.25 GiB 42.66 GiB 4.3 tok/s COMFORTABLE
128k 24.50 GiB 54.91 GiB 3.3 tok/s OOM

Buy-vs-Cloud-vs-API Monthly TCO Break-Even

USE SERVERLESS API
Own Hardware/mo
$48.31
871.7h active + power
Cloud GPU/mo
$892.14
@$0.89/hr spot/pod
Serverless API/mo
$13.50
API cheaper

Recommended Hardware & Cloud Deployment

Contextual to fit status

Hardware FAQ — Ryzen AI Max+ 395 Strix Halo (64 GB)

How much usable VRAM / GPU memory does Ryzen AI Max+ 395 Strix Halo (64 GB) actually provide for LLMs?

Out of the box, Ryzen AI Max+ 395 Strix Halo (64 GB) safely reserves 48 GiB for GPU inference, expandable up to 54 GiB with OS UMA tuning.

What is the largest Q4_K_M model Ryzen AI Max+ 395 Strix Halo (64 GB) can run at 32k context?

With Q4_K_M weights and FP16 KV cache at 32k context, the largest fitting model in our verified catalog on Ryzen AI Max+ 395 Strix Halo (64 GB) is Llama-3.3 Nemotron Super 49B (49.9B total).