AMD Ryzen AI Max+ 395
DesktopA unified-memory APU (codename Strix Halo): up to 128 GB of unified memory, of which up to 96 GB can be allocated as VRAM, at 256 GB/s. Default TDP is 55 W with a configurable 45–120 W range. Ships in mini-desktops such as the Framework Desktop as well as laptops — the x86 answer to Apple's unified memory.
136
Run records
33
Models covered
4
Frameworks covered
L2 Cross-framework verified × 136
Evidence mix
Typical measured performance
Published benchmark averages per model × framework × quantization on this hardware.
- llama.cpp · Q2_K Decode 42.14 · Prefill 173.71 ·
- Ollama · Q3_K_L Decode 32.47 · Prefill 141.58 ·
- llama.cpp · IQ1_S Decode 115.1 · Prefill 632.07 ·
- vLLM · IQ4_XS Decode 84.12 · Prefill 383.03 ·
- llama.cpp · Q4_K_M Decode 35.76 · Prefill 167.23 ·
- Ollama · Q4_K_M Decode 33.77 · Prefill 108.98 ·
- llama.cpp · IQ3_XS Decode 77.09 · Prefill 382.38 ·
- vLLM · Q4_K_S Decode 63.8 · Prefill 213.74 ·
- llama.cpp · IQ2_M Decode 46.26 · Prefill 241.55 ·
- Ollama · IQ2_M Decode 44.23 · Prefill 204.06 ·
- llama.cpp · Q4_K_S Decode 26.2 · Prefill 94.34 ·
- vLLM · Q4_K_S Decode 27.6 · Prefill 131.32 ·
- llama.cpp · Q3_K_L Decode 32.72 · Prefill 137.17 ·
- vLLM · Q6_K Decode 22.65 · Prefill 118.27 ·
- vLLM · Q8_0 Decode 19.62 · Prefill 103.78 ·
- vLLM · FP16 Decode 9.75 · Prefill 49.87 ·
- vLLM · Q8_0 Decode 13.23 · Prefill 57.54 ·
- Ollama · Q3_K_L Decode 19.34 · Prefill 100.78 ·
- llama.cpp · IQ3_XS Decode 27.72 · Prefill 124.12 ·
- Ollama · FP16 Decode 12 · Prefill 64.43 ·
- vLLM · Q3_K_S Decode 72.5 · Prefill 334.68 ·
- vLLM · Q6_K Decode 44.31 · Prefill 168.04 ·
- Ollama · Q5_K_M Decode 45.61 · Prefill 217.15 ·
- vLLM · IQ2_M Decode 61.41 · Prefill 198.56 ·
- Ollama · FP16 Decode 19.79 · Prefill 92.05 ·
- vLLM · Q8_0 Decode 18.22 · Prefill 78.36 ·
- Ollama · IQ3_S Decode 27.26 · Prefill 100.74 ·
- Ollama · Q6_K Decode 16.05 · Prefill 66.18 ·
- llama.cpp · IQ2_XS Decode 231.64 · Prefill 936.81 ·
- llama.cpp · Q4_K_M Decode 68.16 · Prefill 362.8 ·
- vLLM · IQ4_XS Decode 85.62 · Prefill 384.79 ·
- llama.cpp · Q4_K_M Decode 26.52 · Prefill 94.42 ·
- vLLM · IQ2_M Decode 42.37 · Prefill 221.74 ·
- vLLM · Q6_K Decode 24.65 · Prefill 111.72 ·
- vLLM · Q3_K_S Decode 38.79 · Prefill 186.13 ·
- Ollama · FP16 Decode 14.12 · Prefill 69.25 ·
- Ollama · Q3_K_M Decode 34.18 · Prefill 132.79 ·
- llama.cpp · IQ2_M Decode 24.41 · Prefill 110.5 ·
- llama.cpp · IQ2_XS Decode 25.14 · Prefill 89.05 ·
- Ollama · IQ2_M Decode 24.86 · Prefill 123.01 ·
- llama.cpp · IQ2_M Decode 32.57 · Prefill 131.77 ·
- llama.cpp · Q5_K_M Decode 22.71 · Prefill 109.92 ·
- Ollama · Q3_K_M Decode 26.82 · Prefill 137.08 ·
- Ollama · Q5_K_S Decode 20.2 · Prefill 85.22 ·
- llama.cpp · IQ3_XS Decode 45.88 · Prefill 214.9 ·
- vLLM · Q2_K Decode 51.17 · Prefill 219.96 ·
- vLLM · IQ2_M Decode 48.9 · Prefill 230.21 ·
- vLLM · Q5_K_M Decode 31.8 · Prefill 157.2 ·
- Ollama · Q3_K_L Decode 37.4 · Prefill 165.99 ·
- llama.cpp · Q8_0 Decode 61.59 · Prefill 327.72 ·
Fit check
Physics-based estimate from VRAM and memory bandwidth: pick a model to see which quantizations fit and the expected speed
Formula: tok/s = (bandwidth ÷ model size) × 0.55, adapted from llmfit; see run records for measured numbers
Key specs
- Vendor
- AMD
- Architecture
- Zen 5 + RDNA 3.5 (Strix Halo)
- VRAM
- 128 GB
- Memory bandwidth
- 256 GB/s
- TDP
- 55 W
- Release date
- 2025-03-17
Run records
50 configurations
| Model | Framework | Quantization | Decode | Gen time | Samples |
|---|---|---|---|---|---|
| Ministral-3-8B-Instruct-2512 | llama.cpp | Q2_K | 42.1 tok/s | — | 1 |
| Ministral-3-8B-Instruct-2512 | Ollama | Q3_K_L | 32.5 tok/s | — | 1 |
| iFlytek Spark-X2.5-4B | llama.cpp | IQ1_S | 115.1 tok/s | — | 1 |
| iFlytek Spark-X2.5-4B | vLLM | IQ4_XS | 84.1 tok/s | — | 1 |
| NVIDIA-Nemotron-3.5-Bolt-E4B-4GB | llama.cpp | Q4_K_M | 35.8 tok/s | — | 1 |
| NVIDIA-Nemotron-3.5-Bolt-E4B-4GB | Ollama | Q4_K_M | 33.8 tok/s | — | 1 |
| Flint-40B-A2.5B | llama.cpp | IQ3_XS | 77.1 tok/s | — | 1 |
| Flint-40B-A2.5B | vLLM | Q4_K_S | 63.8 tok/s | — | 1 |
| Raon-1-mini-preview | llama.cpp | IQ2_M | 46.3 tok/s | — | 1 |
| Raon-1-mini-preview | Ollama | IQ2_M | 44.2 tok/s | — | 1 |
| mammoth-1 | llama.cpp | Q4_K_S | 26.2 tok/s | — | 1 |
| mammoth-1 | vLLM | Q4_K_S | 27.6 tok/s | — | 1 |
| elephant-1 | llama.cpp | Q3_K_L | 32.7 tok/s | — | 1 |
| elephant-1 | vLLM | Q6_K | 22.6 tok/s | — | 1 |
| elephant-1 | vLLM | Q8_0 | 19.6 tok/s | — | 1 |
| Qwen3.8-27B | vLLM | FP16 | 9.8 tok/s | — | 1 |
| Qwen3.8-27B | vLLM | Q8_0 | 13.2 tok/s | — | 1 |
| Qwen3.8-27B | Ollama | Q3_K_L | 19.3 tok/s | — | 1 |
| Qwen3.6-27B | llama.cpp | IQ3_XS | 27.7 tok/s | — | 1 |
| Qwen3.6-27B | Ollama | FP16 | 12 tok/s | — | 1 |
| Qwen3.6-35B-A3B | vLLM | Q3_K_S | 72.5 tok/s | — | 1 |
| Qwen3.6-35B-A3B | vLLM | Q6_K | 44.3 tok/s | — | 1 |
| Qwen3.6-35B-A3B | Ollama | Q5_K_M | 45.6 tok/s | — | 1 |
| Qwen3.5-4B | vLLM | IQ2_M | 61.4 tok/s | — | 1 |
| Qwen3.5-4B | Ollama | FP16 | 19.8 tok/s | — | 1 |
| Qwen3.5-27B | vLLM | Q8_0 | 18.2 tok/s | — | 1 |
| Qwen3.5-27B | Ollama | IQ3_S | 27.3 tok/s | — | 1 |
| Qwen3.5-27B | Ollama | Q6_K | 16.1 tok/s | — | 1 |
| Qwen3.5-0.8B | llama.cpp | IQ2_XS | 231.6 tok/s | — | 2 |
| Qwen3.5-0.8B | llama.cpp | Q4_K_M | 68.2 tok/s | — | 1 |
| Qwen3.5-0.8B | vLLM | IQ4_XS | 85.6 tok/s | — | 1 |
| Qwen3.8-9B | llama.cpp | Q4_K_M | 26.5 tok/s | — | 1 |
| Qwen3.8-9B | vLLM | IQ2_M | 42.4 tok/s | — | 1 |
| Qwen3.8-9B | vLLM | Q6_K | 24.6 tok/s | — | 1 |
| Qwen3-8B | vLLM | Q3_K_S | 38.8 tok/s | — | 1 |
| Qwen3-8B | Ollama | FP16 | 14.1 tok/s | — | 1 |
| Qwen3-8B | Ollama | Q3_K_M | 34.2 tok/s | — | 1 |
| gemma-4-31B-it | llama.cpp | IQ2_M | 24.4 tok/s | — | 1 |
| gemma-4-31B-it | llama.cpp | IQ2_XS | 25.1 tok/s | — | 1 |
| gemma-4-31B-it | Ollama | IQ2_M | 24.9 tok/s | — | 1 |
| gemma-4-12B-it | llama.cpp | IQ2_M | 32.6 tok/s | — | 1 |
| gemma-4-12B-it | llama.cpp | Q5_K_M | 22.7 tok/s | — | 1 |
| gemma-4-12B-it | Ollama | Q3_K_M | 26.8 tok/s | — | 1 |
| gemma-4-12B-it | Ollama | Q5_K_S | 20.2 tok/s | — | 1 |
| Llama-3.1-8B-Instruct | llama.cpp | IQ3_XS | 45.9 tok/s | — | 1 |
| Llama-3.1-8B-Instruct | vLLM | Q2_K | 51.2 tok/s | — | 1 |
| Llama-3-8B-Instruct | vLLM | IQ2_M | 48.9 tok/s | — | 1 |
| Llama-3-8B-Instruct | vLLM | Q5_K_M | 31.8 tok/s | — | 1 |
| Llama-3-8B-Instruct | Ollama | Q3_K_L | 37.4 tok/s | — | 1 |
| Llama-3.2-3B-Instruct | llama.cpp | Q8_0 | 61.6 tok/s | — | 1 |