← All machines
AMD unified memory · Beelink
Beelink GTR9 Pro (Ryzen AI Max+ 395)
A Strix Halo mini PC with 128GB, an internal power supply and two 10GbE ports.
Memory for models
96 GB
of 128 GB in the largest config
Memory bandwidth
256 GB/s
Maker price
~$4,699
gpt-oss-20b, 4-bit
73 tok/smedian of 3 runs
Memory configurations
One page covers every memory option the maker sells. Prices are the maker’s list price for the cheapest configuration with that much memory.
| Memory | Price | Memory the GPU can use |
|---|---|---|
| 128 GB | ~$4,699 | 96 GB gpu-usable memory |
Windows: AMD Variable Graphics Memory up to 96GB dedicated (112GB total graphics-addressable) on 128GB Ryzen AI Max+ systems. Linux: GTT defaults to ~50% of RAM and can be raised with the ttm pages_limit setting (AMD shows 100GB on a 128GB system). How GPU memory limits work on unified-memory machines.
What it runs at 4-bit
Every model in the catalog at its 4-bit file with 8,192 tokens of context. 29 of 41 fit. Speeds are the median of measured runs on the same chip where any exist; otherwise an estimate from memory bandwidth. Try other quants and context lengths.
| Model | Needs | Fit | Generation speed |
|---|---|---|---|
| Kimi K32.8T, 104B active · UD-Q4_K_XL | 1406.8 GBcontext memory estimated | Does not fit | |
| DeepSeek V4 Pro1.6T, 49B active · Q4_K_M | 890.5 GBcontext memory not counted | Does not fit | |
| Kimi K2.61T, 32B active · UD-Q4_K_XL | 545.7 GBcontext memory estimated | Does not fit | |
| DeepSeek V4.1 Flash552B, 16B active · Q4 (DwarfStar) | 484.5 GB | Does not fit | |
| GLM-5.3753.3B · UD-Q4_K_XL | 437.4 GBcontext memory estimated | Does not fit | |
| DeepSeek V3.2671B, 37B active · Q4_K_M | 379.6 GBcontext memory estimated | Does not fit | |
| MiniMax M3428B, 23B active · Q4_K_M | 244.8 GBcontext memory not counted | Does not fit | |
| Qwen3.5 397B-A17B397B, 17B active · Q4_K_M | 229.1 GBcontext memory estimated | Does not fit | |
| GLM-5.3 Flash320B, 18B active · UD-Q4_K_XL | 187.6 GBcontext memory estimated | Does not fit | |
| Hy3295B, 21B active · Q4_K_M | 173.6 GB | Does not fit | |
| MiniMax M2.7228.7B · Q4_K_M | 132.7 GB | Does not fit | |
| Qwen3.8 Flash-Next125B, 6B active · Q4_K_M | 113.1 GBcontext memory estimated | Does not fit | |
| Nemotron 3 Super 120B-A12B120B, 12B active · Q4_K_M | 82.6 GBcontext memory estimated | Fits | 15 tok/smedian of 2 runs |
| Qwen3.5 122B-A10B122B, 10B active · Q4_K_M | 73.0 GBcontext memory estimated | Fits | 20 tok/smedian of 2 runs |
| Mistral Medium 3.5 128B128B · Q4_K_M | 74.0 GB | Fits | ~2.3–2.8 tok/sestimate |
| GLM-4.5-Air106B, 12B active · Q4_K_M | 70.9 GB | Fits | ~15–19 tok/sestimate |
| Mistral Small 4 119B (2603)119B, 6.5B active · Q4_K_M | 69.3 GBcontext memory estimated | Fits | ~30–39 tok/sestimate |
| Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M | 62.4 GBcontext memory not counted | Fits | 18 tok/smedian of 2 runs |
| gpt-oss-120b117B, 5.1B active · MXFP4 (native) | 60.2 GBcontext memory estimated | Fits | 52 tok/smedian of 5 runs |
| Qwen3 Coder Next80B, 3B active · Q4_K_M | 46.9 GBcontext memory estimated | Fits | ~66–85 tok/sestimate |
| Llama 3.3 70B Instruct70B · Q4_K_M | 43.6 GB | Fits | 4.8 tok/smedian of 4 runs |
| Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M | 25.3 GBcontext memory estimated | Fits | ~47–60 tok/sestimate |
| Qwen3.6 35B-A3B35B, 3B active · Q4_K_M | 22.4 GBcontext memory estimated | Fits | 60 tok/smedian of 7 runs |
| Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M | 21.9 GBcontext memory estimated | Fits | ~65–83 tok/sestimate |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M | 19.5 GB | Fits | ~60–77 tok/sestimate |
| Gemma 4 31B IT30.7B · Q4_K_M | 18.9 GBcontext memory estimated | Fits | 11 tok/smedian of 2 runs |
| Granite 4.2 30B29.3B · Q4_K_M | 20.3 GB | Fits | ~9.6–12 tok/sestimate |
| Qwen3.8 27B27B · Q4_K_M | 18.2 GBcontext memory estimated | Fits | 12 tok/smedian of 2 runs |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M | 17.4 GBcontext memory estimated | Fits | 52 tok/smedian of 2 runs |
| Qwen3.6 27B27B · Q4_K_M | 17.7 GBcontext memory estimated | Fits | ~10–12 tok/sestimate |
| Devstral Small 2 24B (2512)24B · Q4_K_M | 16.1 GB | Fits | ~12–15 tok/sestimate |
| gpt-oss-20b21B, 3.6B active · MXFP4 (native) | 12.5 GBcontext memory estimated | Fits | 73 tok/smedian of 3 runs |
| Phi-414.7B · Q4_K_M | 11.3 GB | Fits | ~20–23 tok/sestimate |
| Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M | 10.4 GB | Fits | ~21–25 tok/sestimate |
| Gemma 4 12B IT11.95B · Q4_K_M | 8.2 GBcontext memory estimated | Fits | ~24–29 tok/sestimate |
| Qwen3.5 9B9B · Q4_K_M | 7.0 GBcontext memory estimated | Fits | ~31–37 tok/sestimate |
| Granite 4.2 8B8.8B · Q4_K_M | 7.9 GB | Fits | ~31–38 tok/sestimate |
| Gemma 4 E4B IT8B · Q4_K_M | 6.5 GBcontext memory not counted | Fits | ~57–68 tok/sestimate |
| Qwen3 8B8.2B · Q4_K_M | 7.3 GB | Fits | ~35–41 tok/sestimate |
| Llama 3.1 8B Instruct8B · Q4_K_M | 7.1 GB | Fits | 40 tok/smedian of 2 runs |
| Qwen3.5 4B4B · Q4_K_M | 4.3 GBcontext memory estimated | Fits | ~63–76 tok/sestimate |
Measured speeds
78 published runs with one user, plain decoding and at most 4,096 tokens already in context, on Ryzen AI Max+ 395. Runs on 5 other machines with the same chip count here too, and say so.
| Model | Quant | Backend | Generation | Prompt | Context | Measured on | Source |
|---|---|---|---|---|---|---|---|
| Gemma 4 26B-A4B IT | BF16 | llama.cpp b9187 (ROCm 7.2.3) | 25 tok/s | 909 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | BF16 | llama.cpp b9187 (Vulkan RADV) | 19 tok/s | 588 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q4_K_XL | llama.cpp b9187 (Vulkan RADV) | 55 tok/s | 1,325 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q4_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 48 tok/s | 1,329 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q8_K_XL | llama.cpp b9187 (Vulkan RADV) | 45 tok/s | 1,291 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q8_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 42 tok/s | 1,354 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | UD-Q8_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 42 tok/s | 1,700 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-07 |
| Gemma 4 31B IT | BF16 | llama.cpp b9187 (ROCm 7.2.3) | 3.5 tok/s | 428 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | BF16 | llama.cpp b9187 (Vulkan RADV) | 3.5 tok/s | 136 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q4_K_XL | llama.cpp b9187 (Vulkan RADV) | 11 tok/s | 302 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q4_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 11 tok/s | 323 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q8_K_XL | llama.cpp b9187 (Vulkan RADV) | 6.3 tok/s | 272 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q8_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 6.2 tok/s | 326 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-120b | MXFP4 | llama.cpp b9187 (Vulkan RADV) | 57 tok/s | 720 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-120b | MXFP4 | llama.cpp b9187 (ROCm 7.2.3) | 52 tok/s | 626 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-120b | MXFP4 | llama.cpp b7123 (ROCm 6.4.4) | 52 tok/s | 781 tok/s512-token prompt | 0 | a machine with the same chip | llama.cpp benchmark discussion2025-11 |
| gpt-oss-120b | MXFP4 MoE | llama.cpp b6816 ROCm | 47 tok/s | 1,000 tok/s2,048-token prompt | 0 | GMKtec EVO-X2 (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2025-10 |
| gpt-oss-120b | MXFP4 MoE | llama.cpp b6816 ROCm | 44 tok/s | 824 tok/s2,048-token prompt | 4,096 | GMKtec EVO-X2 (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2025-10 |
| gpt-oss-120b | UD-Q6_K_XL | llama.cpp b6697 (Vulkan RADV) | 54 tok/s | 509 tok/s2,048-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2025-10 |
| gpt-oss-120b | Q6_K | llama.cpp (build not stated) Vulkan | 53 tok/s | 504 tok/s2,048-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2025-10 |
| gpt-oss-20b | BF16 | llama.cpp (build not stated) Vulkan | 48 tok/s | 982 tok/s2,048-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2025-10 |
| gpt-oss-20b | MXFP4 | llama.cpp b9187 (Vulkan RADV) | 80 tok/s | 1,692 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-20b | MXFP4 | llama.cpp b9187 (ROCm 7.2.3) | 73 tok/s | 1,786 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-20b | MXFP4 | Ollama (version not stated) | 49 tok/s | – | 0 | Minisforum MS-S1 MAX (Ryzen AI Max+ 395) (same chip) | review2026-03 |
| Llama 3.1 8B Instruct | i1-Q4_K_M | llama.cpp ~b5863 (Vulkan RADV, fa=1) | 42 tok/s | 614 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | owner report2025-07 |
| Llama 3.1 8B Instruct | i1-Q4_K_M | llama.cpp ~b5863 (HIP hipBLASLt) | 37 tok/s | 878 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | owner report2025-07 |
| Llama 3.3 70B Instruct | i1-Q4_K_M | llama.cpp b5349 (Vulkan RADV) | 5.0 tok/s | 77 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2025-08 |
| Llama 3.3 70B Instruct | i1-Q4_K_M | llama.cpp ~b5863 (Vulkan RADV, fa=1) | 5.0 tok/s | 26 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | owner report2025-07 |
| Llama 3.3 70B Instruct | i1-Q4_K_M | llama.cpp b5348 (ROCm/HIP rocWMMA) | 4.7 tok/s | 34 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2025-08 |
| Llama 3.3 70B Instruct | i1-Q4_K_M | llama.cpp ~b5863 (HIP rocWMMA) | 4.5 tok/s | 95 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | owner report2025-07 |
| Llama 4 Scout 17B-16E Instruct | UD-Q4_K_XL | llama.cpp ~b5863 (Vulkan RADV, fa=1 b=256) | 19 tok/s | 146 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | owner report2025-07 |
| Llama 4 Scout 17B-16E Instruct | UD-Q4_K_XL | llama.cpp ~b5863 (HIP hipBLASLt) | 17 tok/s | 264 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | owner report2025-07 |
| MiniMax M2.7 | Q3_K_S | llama.cpp b9187 (Vulkan RADV) | 31 tok/s | 242 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| MiniMax M2.7 | Q3_K_S | llama.cpp b9187 (ROCm 7.2.3) | 23 tok/s | 243 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Nemotron 3 Super 120B-A12B | Q4_K_XL | llama.cpp b9193 (Vulkan RADV) | 15 tok/s | 271 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Nemotron 3 Super 120B-A12B | Q4_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 15 tok/s | 276 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3 Coder Next | not stated (Ollama default tag, 51.7 GB) | Ollama (version not stated) | 30 tok/s | – | 0 | Minisforum MS-S1 MAX (Ryzen AI Max+ 395) (same chip) | review2026-03 |
| Qwen3.5 122B-A10B | UD-Q4_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 22 tok/s | 487 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.5 122B-A10B | Q4_K_M | Ollama (version not stated) | 19 tok/s | – | 0 | Minisforum MS-S1 MAX (Ryzen AI Max+ 395) (same chip) | review2026-03 |
| Qwen3.5 122B-A10B | Q5_K_XL | llama.cpp b9193 (Vulkan RADV) | 22 tok/s | 309 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.5 122B-A10B | Q5_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 20 tok/s | 337 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 27B | Q8_0 | llama.cpp b9927 (ROCm 7.2.4) | 7.8 tok/s | 336 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 27B | Q8_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 6.5 tok/s | 367 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 27B | Q8_K_XL | llama.cpp b9193 (Vulkan RADV) | 6.3 tok/s | 275 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | BF16 | llama.cpp b9193 (ROCm 7.2.3) | 26 tok/s | 526 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | BF16 | llama.cpp b9193 (Vulkan RADV) | 11 tok/s | 328 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q4_K_M | llama.cpp b9870 (Vulkan RADV) | 61 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-07 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp performance fork (build not stated) (Vulkan RADV) | 61 tok/s | 1,750 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (build not stated) (Vulkan RADV) | 61 tok/s | 1,219 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | Q4_K_XL | llama.cpp b9193 (Vulkan RADV) | 60 tok/s | 1,114 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 52 tok/s | 1,384 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 35B-A3B | Q4_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 52 tok/s | 1,120 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 51 tok/s | 1,382 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | APEX I-Balanced (Q5_K_M) | llama.cpp master 2026-08 (Vulkan RADV) | 67 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.6 35B-A3B | APEX I-Balanced (Q5_K_M) | llama.cpp master 2026-08 (ROCm HIP) | 56 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp (build not stated) (Vulkan RADV) | 48 tok/s | 1,226 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp performance fork (build not stated) (Vulkan RADV) | 47 tok/s | 1,724 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 47 tok/s | 1,215 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 35B-A3B | Q8_K_XL | llama.cpp b9193 (Vulkan RADV) | 46 tok/s | 1,045 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | Q8_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 46 tok/s | 1,094 tok/s512-token prompt | 0 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 46 tok/s | 1,243 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 27B | UD-Q4_K_XL | Gufo 0.9.1 (ROCm 10.0) | 12 tok/s | 751 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-10 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 12 tok/s | 386 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp b10435-era (9d57ce4, ROCm/HIP) | – | 390 tok/s16,000-token prompt | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q5_K_M | llama.cpp master 2026-08 (Vulkan RADV) | 11 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q5_K_M | llama.cpp master 2026-08 (ROCm HIP) | 11 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | UD-Q8_K_XL | llama.cpp (official server-rocm Docker image 2026-08-16, ROCm) | – | 80 tok/s131,072-token prompt | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | UD-Q8_K_XL | llama.cpp (Vulkan build, version not stated) | – | 40 tok/s131,072-token prompt | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 Flash-Next | Q2 | DwarfStar (ROCm 10.0) | 21 tok/s | 565 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ3_XXS | llama.cpp performance fork (build not stated) (Vulkan RADV) | 29 tok/s | 583 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ3_XXS | llama.cpp (build not stated) (ROCm 10.0) | 22 tok/s | 366 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp performance fork (build not stated) (Vulkan RADV) | 26 tok/s | 599 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-Q4_K_XL | Gufo 0.9.1 (ROCm 10.0) | 26 tok/s | 1,681 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-10 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp (build not stated) (ROCm 10.0) | 21 tok/s | 470 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | Q4 | DwarfStar (ROCm 10.0) | 20 tok/s | 505 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | 33 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | – | 510 tok/s8,000-token prompt | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | – | 390 tok/s32,000-token prompt | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
With speculative decoding
A small draft model or extra prediction heads guess several tokens ahead. Generation is faster on text the guesses get right, so these are kept out of the medians above.
| Model | Quant | Backend | Generation | Prompt | Context | Measured on | Source |
|---|---|---|---|---|---|---|---|
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (build not stated) (Vulkan RADV) | 89 tok/s | 1,167 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q4_K_M (MTP variant) | llama.cpp b9870 (Vulkan RADV) | 85 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-07 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp performance fork (build not stated) (Vulkan RADV) | 78 tok/s | 1,634 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 68 tok/s | 1,325 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | APEX I-Balanced (Q5_K_M) | llama.cpp master 2026-08 (Vulkan RADV) | 81 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp (build not stated) (Vulkan RADV) | 67 tok/s | 1,179 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 65 tok/s | 1,225 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp performance fork (build not stated) (Vulkan RADV) | 58 tok/s | 1,425 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp b10435-era (9d57ce4, ROCm/HIP) | 60 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | UD-Q4_K_XL | Gufo 0.9.1 (ROCm 10.0) | 33 tok/s | 677 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-10 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 22 tok/s | 356 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp fork (dflash tree 9d57ce4, ROCm 7.2.4) | – | 317 tok/s6,000-token prompt | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp fork (dflash tree 9d57ce4, ROCm 7.2.4) | – | 168 tok/s35,000-token prompt | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 27B | Q5_K_M | llama.cpp master 2026-08 (Vulkan RADV) | 17 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 Flash-Next | Q2 | DwarfStar (ROCm 10.0) | 23 tok/s | 564 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ3_XXS | EngramHalo SSD (ROCm 10.0) | 24 tok/s | 450 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ3_XXS | EngramHalo RAM (ROCm 10.0) | 21 tok/s | 455 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-Q4_K_XL | Gufo 0.9.1 (ROCm 10.0) | 48 tok/s | 1,620 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-10 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp fork (drluoto/strix-halo-flash-next, ROCm 7.1 HIP) | 47 tok/s | – | 8,000 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 Flash-Next | Q4_0_ROCMFP4_FAST | llama.cpp fork (LaurentZuijdwijk qwen4exp MTP-draft merged with upstream, b10981, Vulkan)2 requests at once | 47 tok/s | 192 tok/s24,000-token prompt | 24,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp fork (stew675 rdna-boosts, HIP via TheRock 10.1 SDK, Lemonade, Windows 11) | 38 tok/s | – | – | Corsair AI Workstation 300 (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp (drluoto strix-halo-vulkan fork; build not stated) (Vulkan RADV) | 38 tok/s | 390 tok/s | 32,768 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan) | 33 tok/s | – | – | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp fork (drluoto/strix-halo-flash-next, ROCm 7.1 HIP) | 32 tok/s | – | 8,000 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan) | 27 tok/s | – | – | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp fork (drluoto/strix-halo-flash-next, ROCm 7.1 HIP) | 25 tok/s | – | 24,000 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 Flash-Next | Q4 | DwarfStar (ROCm 10.0) | 24 tok/s | 504 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | EngramHalo RAM (ROCm 10.0) | 23 tok/s | 464 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | EngramHalo SSD (ROCm 10.0) | 18 tok/s | 467 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant (drluoto AgenticRequant-Q5K) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan) | 61 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | 58 tok/s | – | 0 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant (drluoto AgenticRequant-Q5K) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan) | 43 tok/s | – | 8,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | 42 tok/s | – | 8,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant (drluoto AgenticRequant-Q5K) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan) | 39 tok/s | – | – | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant (drluoto AgenticRequant-Q5K) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan) | 38 tok/s | – | 32,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | 38 tok/s | – | 32,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant (drluoto AgenticRequant-Q5K) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan) | 32 tok/s | – | 8,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | 30 tok/s | – | 8,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant (drluoto AgenticRequant-Q5K) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan) | 29 tok/s | – | – | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | W4B + Q8 overlay | Halogen (ROCm external) | 36 tok/s | 822 tok/s2,048-token prompt | 0 | a machine with the same chip | independent benchmark2026-09 |
Deep into a long context
Measured with more than 4,096 tokens already in context, where every new token also reads the context memory.
| Model | Quant | Backend | Generation | Prompt | Context | Measured on | Source |
|---|---|---|---|---|---|---|---|
| Gemma 4 26B-A4B IT | BF16 | llama.cpp b9187 (ROCm 7.2.3) | 22 tok/s | 667 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | BF16 | llama.cpp b9187 (ROCm 7.2.3) | 21 tok/s | 447 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | BF16 | llama.cpp b9187 (Vulkan RADV) | 17 tok/s | 428 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | BF16 | llama.cpp b9187 (Vulkan RADV) | 16 tok/s | 332 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q4_K_XL | llama.cpp b9187 (Vulkan RADV) | 42 tok/s | 682 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q4_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 38 tok/s | 725 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q4_K_XL | llama.cpp b9187 (Vulkan RADV) | 37 tok/s | 470 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q4_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 36 tok/s | 471 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q8_K_XL | llama.cpp b9187 (Vulkan RADV) | 37 tok/s | 673 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | UD-Q8_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 36 tok/s | 951 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-07 |
| Gemma 4 26B-A4B IT | UD-Q8_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 34 tok/s | 651 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-07 |
| Gemma 4 26B-A4B IT | Q8_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 34 tok/s | 726 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q8_K_XL | llama.cpp b9187 (Vulkan RADV) | 33 tok/s | 466 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 26B-A4B IT | Q8_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 32 tok/s | 472 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | BF16 | llama.cpp b9187 (Vulkan RADV) | 3.3 tok/s | 92 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | BF16 | llama.cpp b9187 (ROCm 7.2.3) | 3.2 tok/s | 167 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | BF16 | llama.cpp b9187 (ROCm 7.2.3) | 3.1 tok/s | 109 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | BF16 | llama.cpp b9187 (Vulkan RADV) | 3.1 tok/s | 68 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q4_K_XL | llama.cpp b9187 (Vulkan RADV) | 9.1 tok/s | 146 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q4_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 8.1 tok/s | 150 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q4_K_XL | llama.cpp b9187 (Vulkan RADV) | 7.9 tok/s | 87 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q4_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 7.5 tok/s | 102 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q8_K_XL | llama.cpp b9187 (Vulkan RADV) | 5.5 tok/s | 141 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q8_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 5.3 tok/s | 150 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q8_K_XL | llama.cpp b9187 (Vulkan RADV) | 5.0 tok/s | 88 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Gemma 4 31B IT | Q8_K_XL | llama.cpp b9187 (ROCm 7.2.3) | 5.0 tok/s | 101 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-120b | MXFP4 | llama.cpp b9187 (Vulkan RADV) | 43 tok/s | 308 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-120b | MXFP4 MoE | llama.cpp b6816 ROCm | 43 tok/s | 703 tok/s2,048-token prompt | 8,192 | GMKtec EVO-X2 (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2025-10 |
| gpt-oss-120b | MXFP4 MoE | llama.cpp b6816 ROCm | 40 tok/s | 515 tok/s2,048-token prompt | 16,384 | GMKtec EVO-X2 (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2025-10 |
| gpt-oss-120b | MXFP4 | llama.cpp b9187 (ROCm 7.2.3) | 36 tok/s | 589 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-120b | MXFP4 MoE | llama.cpp b6816 ROCm | 35 tok/s | 349 tok/s2,048-token prompt | 32,768 | GMKtec EVO-X2 (Ryzen AI Max+ 395) (same chip) | llama.cpp benchmark discussion2025-10 |
| gpt-oss-120b | MXFP4 | llama.cpp b9187 (Vulkan RADV) | 35 tok/s | 146 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-120b | MXFP4 | llama.cpp b9187 (ROCm 7.2.3) | 26 tok/s | 371 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-20b | MXFP4 | llama.cpp b9187 (Vulkan RADV) | 61 tok/s | 556 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-20b | MXFP4 | llama.cpp b9187 (ROCm 7.2.3) | 52 tok/s | 1,003 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-20b | MXFP4 | llama.cpp b9187 (Vulkan RADV) | 50 tok/s | 238 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| gpt-oss-20b | MXFP4 | llama.cpp b9187 (ROCm 7.2.3) | 41 tok/s | 602 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| MiniMax M2.7 | Q3_K_S | llama.cpp b9187 (Vulkan RADV) | 15 tok/s | 34 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| MiniMax M2.7 | Q3_K_S | llama.cpp b9187 (Vulkan RADV) | 8.8 tok/s | 14 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| MiniMax M2.7 | Q3_K_S | llama.cpp b9187 (ROCm 7.2.3) | 6.2 tok/s | 143 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Nemotron 3 Super 120B-A12B | Q4_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 14 tok/s | 382 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Nemotron 3 Super 120B-A12B | Q4_K_XL | llama.cpp b9193 (Vulkan RADV) | 14 tok/s | 243 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Nemotron 3 Super 120B-A12B | Q4_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 14 tok/s | 348 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Nemotron 3 Super 120B-A12B | Q4_K_XL | llama.cpp b9193 (Vulkan RADV) | 14 tok/s | 216 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.5 122B-A10B | UD-Q4_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 20 tok/s | 330 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.5 122B-A10B | UD-Q4_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 18 tok/s | 271 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.5 122B-A10B | Q5_K_XL | llama.cpp b9193 (Vulkan RADV) | 20 tok/s | 225 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.5 122B-A10B | Q5_K_XL | llama.cpp b9193 (Vulkan RADV) | 19 tok/s | 165 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.5 122B-A10B | Q5_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 18 tok/s | 276 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.5 122B-A10B | Q5_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 16 tok/s | 192 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 27B | Q8_0 | llama.cpp b9927 (ROCm 7.2.4) | 7.5 tok/s | 246 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 27B | Q8_0 | llama.cpp b9927 (ROCm 7.2.4) | 7.3 tok/s | 198 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 27B | Q8_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 6.1 tok/s | 198 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 27B | Q8_K_XL | llama.cpp b9193 (Vulkan RADV) | 5.9 tok/s | 123 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 27B | Q8_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 5.8 tok/s | 106 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 27B | Q8_K_XL | llama.cpp b9193 (Vulkan RADV) | 5.6 tok/s | 63 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | BF16 | llama.cpp b9193 (ROCm 7.2.3) | 24 tok/s | 418 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | BF16 | llama.cpp b9193 (ROCm 7.2.3) | 22 tok/s | 323 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | BF16 | llama.cpp b9193 (Vulkan RADV) | 10 tok/s | 276 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | BF16 | llama.cpp b9193 (Vulkan RADV) | 9.9 tok/s | 241 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp performance fork (build not stated) (Vulkan RADV) | 55 tok/s | 1,261 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (build not stated) (Vulkan RADV) | 54 tok/s | 920 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp performance fork (build not stated) (Vulkan RADV) | 51 tok/s | 971 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (build not stated) (Vulkan RADV) | 50 tok/s | 728 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | Q4_K_XL | llama.cpp b9193 (Vulkan RADV) | 49 tok/s | 716 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 48 tok/s | 895 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 46 tok/s | 894 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 44 tok/s | 677 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | Q4_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 44 tok/s | 747 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | Q4_K_XL | llama.cpp b9193 (Vulkan RADV) | 43 tok/s | 525 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 42 tok/s | 661 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 35B-A3B | Q4_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 38 tok/s | 493 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp (build not stated) (Vulkan RADV) | 44 tok/s | 902 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp performance fork (build not stated) (Vulkan RADV) | 44 tok/s | 1,225 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 43 tok/s | 837 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 42 tok/s | 909 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp performance fork (build not stated) (Vulkan RADV) | 41 tok/s | 968 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp (build not stated) (Vulkan RADV) | 41 tok/s | 722 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 41 tok/s | 638 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.6 35B-A3B | Q8_K_XL | llama.cpp b9193 (Vulkan RADV) | 40 tok/s | 683 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | Q8_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 40 tok/s | 729 tok/s2,048-token prompt | 32,768 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | UD-Q8_K_XL | llama.cpp b9927 (ROCm 7.2.4) | 39 tok/s | 649 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-07 |
| Qwen3.6 35B-A3B | Q8_K_XL | llama.cpp b9193 (Vulkan RADV) | 36 tok/s | 505 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.6 35B-A3B | Q8_K_XL | llama.cpp b9193 (ROCm 7.2.3) | 35 tok/s | 486 tok/s2,048-token prompt | 65,536 | Framework Desktop (Ryzen AI Max+ 395) (same chip) | independent benchmark2026-05 |
| Qwen3.8 27B | UD-Q4_K_XL | Gufo 0.9.1 (ROCm 10.0) | 11 tok/s | 625 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-10 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 11 tok/s | 285 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 27B | UD-Q4_K_XL | Gufo 0.9.1 (ROCm 10.0) | 11 tok/s | 543 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-10 |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp (build not stated) (ROCm 10.0) | 10 tok/s | 174 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | Q2 | DwarfStar (ROCm 10.0) | 21 tok/s | 457 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | Q2 | DwarfStar (ROCm 10.0) | 21 tok/s | 482 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ3_XXS | llama.cpp performance fork (build not stated) (Vulkan RADV) | 24 tok/s | 451 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ3_XXS | llama.cpp performance fork (build not stated) (Vulkan RADV) | 21 tok/s | 367 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ3_XXS | llama.cpp (build not stated) (ROCm 10.0) | 20 tok/s | 330 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ3_XXS | llama.cpp (build not stated) (ROCm 10.0) | 17 tok/s | 269 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp (drluoto strix-halo-vulkan fork; build not stated) (Vulkan RADV) | 30 tok/s | – | 8,192 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | UD-Q4_K_XL | Gufo 0.9.1 (ROCm 10.0) | 26 tok/s | 1,464 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-10 |
| Qwen3.8 Flash-Next | UD-Q4_K_XL | Gufo 0.9.1 (ROCm 10.0) | 26 tok/s | 1,358 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-10 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp (drluoto strix-halo-vulkan fork; build not stated) (Vulkan RADV) | 25 tok/s | – | 32,768 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp performance fork (build not stated) (Vulkan RADV) | 23 tok/s | 462 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp performance fork (build not stated) (Vulkan RADV) | 20 tok/s | 373 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | Q4 | DwarfStar (ROCm 10.0) | 20 tok/s | 437 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | Q4 | DwarfStar (ROCm 10.0) | 19 tok/s | 415 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp (build not stated) (ROCm 10.0) | 18 tok/s | 334 tok/s2,048-token prompt | 16,384 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp fork (drluoto/strix-halo-flash-next, ROCm 7.1 HIP) | 17 tok/s | – | 8,000 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp (build not stated) (ROCm 10.0) | 17 tok/s | 275 tok/s2,048-token prompt | 32,768 | a machine with the same chip | independent benchmark2026-09 |
| Qwen3.8 Flash-Next | UD-IQ4_XS | llama.cpp fork (drluoto/strix-halo-flash-next, ROCm 7.1 HIP) | 15 tok/s | – | 24,000 | a machine with the same chip | llama.cpp benchmark discussion2026-08 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | 30 tok/s | – | 8,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
| Qwen3.8 Flash-Next | Q5_K requant of UD-IQ4_XS (drluoto AgenticRequant-Q5K; per #27950) | llama.cpp fork (drluoto/strix-halo-vulkan, Vulkan RADV) | 25 tok/s | – | 32,000 | a machine with the same chip | llama.cpp benchmark discussion2026-09 |
Specifications
- Chip
- Ryzen AI Max+ 395
- Memory
- LPDDR5X-8000, soldered
- Memory bandwidth
- 256 GB/s
- FP16 compute
- AMD quotes up to 126 overall TOPS (CPU+GPU+NPU, precision not stated) and no FP16 TFLOPS figure
- NPU
- 50 TOPS
- Software
- ROCm, Vulkan, DirectML
- Ships with
- Windows 11 Pro
- Power
- 230 W (Built-in 230W PSU (Beelink))
- Form factor
- Mini PC
- Networking
- Dual 10GbE; two USB4 (40Gbps), which Beelink pitches for AI server clusters
- Maker page
- bee-link.com
Sources
- bee-link.com: variant and price, 230W built-in PSU, dual 10GbE, USB4, Windows 11 Pro
- amd.com: 256-bit LPDDR5x-8000, max 128GB, Radeon 8060S 40 CU, NPU 50 TOPS, 126 overall TOPS, cTDP 45-120W
- amd.com: 128GB Ryzen AI Max+ systems: up to 96GB Variable Graphics Memory (dedicated), 112GB total graphics-addressable on Windows
- rocm.docs.amd.com: Linux GTT default ~50% of RAM; raise via ttm pages_limit / amd-ttm --set (AMD example: 100GB on a 128GB system); keep BIOS VRAM small