Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s
| GPU(s) | AMD Ryzen AI Max+ PRO 495 (Radeon 8065S iGPU, 40 CUs) Amazon · eBay (used) · Newegg |
|---|---|
| VRAM / unified memory (GB) | 192 |
| CPU | AMD Ryzen AI Max+ PRO 495 ("Gorgon Halo") Amazon · eBay (used) · Newegg |
| Motherboard | |
| RAM | 192 GB LPDDR5X-8533 unified (GPU limit raised to 170 GB via ttm.pages_limit) |
| PSU | |
| Other | 2 TB SSD, dual 10GbE, Ubuntu Server 26.04.1 (kernel 7.0), llama.cpp Vulkan, flash attention on, q8_0 KV cache, ubatch 2048 |
| Reported total cost | $7,399 (≈€6,602) |
| Power |
Measured speed (generation tokens/sec)
📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
| Model | Quant | Engine | tok/s | Details |
|---|---|---|---|---|
| gpt-oss 120B | MXFP4 | llama.cpp Vulkan | 55.8 | |
| gpt-oss 120B | MXFP4, 32K context | llama.cpp Vulkan | 45.1 | |
| MiniMax M2.7 | IQ4_XS | llama.cpp Vulkan | 29.5 | |
| MiniMax M2.7 | IQ4_XS, 32K context | llama.cpp Vulkan | 17.3 | |
| DeepSeek V4 Flash | IQ3_S | llama.cpp Vulkan | 17.4 | |
| DeepSeek V4 Flash | IQ3_S, 32K context | llama.cpp Vulkan | 12.2 | |
| Qwen3.8 27B | Q8 (author: "about 8 tok/s") | llama.cpp Vulkan | 8 |
Jake's Garage, 2026. Cost is the author's pre-order price. One of the first owner-measured Ryzen AI Max+ 495 results. The author found the default f16 KV cache was the long-context bottleneck on Vulkan; q8_0 KV roughly doubled 32K prompt processing. ROCm was slower than tuned Vulkan for generation.
Source: https://www.hommie.net/posts/ai-box.html
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)