AMD · RDNA 3.5 (Strix Halo) · 2025
AMD Ryzen AI Max+ 395 (128GB) for local AI
The cheapest way to run 100B-class MoE models in one quiet box, though prices have climbed past $3,300.
Updated · Speeds marked ~ are estimates; seemethodology
- Unified memory
- 128GB
- Bandwidth
- 256 GB/s
- Power
- 120W
- Runs well
- 12 models
Ryzen AI Max+ 395
Typically about $3,300
Can the Ryzen AI Max+ 395 run LLMs?
The Ryzen AI Max+ 395 has 128GB of unified memory (about 96GB usable by the GPU by default) and 256 GB/s of memory bandwidth, which decides how big a model fits and how fast it generates text. It comfortably runs models up to about 122B parameters: the largest in our catalog is Qwen3.5 122B-A10B at Q4_K_M (77GB with 8K context, ~18 tokens/sec).
Recommendations
Best models for the Ryzen AI Max+ 395
Best overall · also coding, reasoning, vision, largest that fits
Qwen3.5 122B-A10BQ4_K_M · 77GB · ~18 tokens/sec
Runs wellRunner-up
Qwen3.6 35B-A3BQ8_0 · 41GB · ~30 tokens/sec
Runs wellFastest good model
gpt-oss-20bMXFP4 · 15GB · ~41 tokens/sec
Runs well
Compatibility
What runs on the Ryzen AI Max+ 395
Every current model in our catalog, at the best quantization that fits in the ~96GB the GPU can use with 8K tokens of context. "Partial offload" means part of the model runs from system RAM (we assume 64GB), which is much slower.
| Model | Parameters | Best quant | Memory | Speed | Verdict |
|---|---|---|---|---|---|
| Gemma 4 E4B | 8B | Q8_0 | 10 GB | ~20 tok/s | Runs well |
| Qwen3.5 9B | 9B | Q8_0 | 12 GB | ~17 tok/s | Runs well |
| Phi-4 | 14.7B | Q8_0 | 18 GB | ~12 tok/s | Runs well |
| gpt-oss-20b | 20.9B (3.6B active) | MXFP4 | 15 GB | ~41 tok/s | Runs well |
| Gemma 4 26B A4B | 26B (3.8B active) | Q8_0 | 30 GB | ~26 tok/s | Runs well |
| Qwen3.8 27B | 27B | Q8_0 | 32 GB | ~6.4 tok/s | Runs well |
| Gemma 4 31B | 31B | Q8_0 | 36 GB | ~5.8 tok/s | Runs well |
| Qwen3.6 35B-A3B | 35B (3B active) | Q8_0 | 41 GB | ~30 tok/s | Runs well |
| GLM-4.5-Air | 106B (12B active) | Q5_K_M | 80 GB | ~13 tok/s | Runs well |
| Llama 4 Scout | 109B (17B active) | Q5_K_M | 81 GB | ~9.3 tok/s | Runs well |
| gpt-oss-120b | 117B (5.1B active) | MXFP4 | 68 GB | 47 tok/s | Runs well |
| Qwen3.5 122B-A10B | 122B (10B active) | Q4_K_M | 77 GB | ~18 tok/s | Runs well |
| DeepSeek V4 Flash | 284B (13B active) | — | — | — | Won't fitWhat it needs |
| GLM-5.3-Flash | 320B (18B active) | — | — | — | Won't fitWhat it needs |
| GLM-5.3 | 744B (40B active) | — | — | — | Won't fitWhat it needs |
Performance
Ryzen AI Max+ 395 vs. alternatives on Qwen3.8 27B
Generation speed in tokens per second for a single user. Token generation is limited mostly by memory bandwidth, which is why bandwidth predicts these results so well.
M5 Max 128GB
Q8_0 · 128GB
~13tok/s
DGX Spark
Q8_0 · 128GB
~6.9tok/s
Ryzen AI Max+ PRO 495
Q8_0 · 192GB
~6.9tok/s
Ryzen AI Max+ 395
Q8_0 · 128GB
~6.4tok/s
Measured
Ryzen AI Max+ 395 benchmark results
| Model | Quant | Runtime | Generation | Prompt | Source |
|---|---|---|---|---|---|
| Llama 2 7B Chat | Q4_0 | llama.cpp (Vulkan) | 52 tok/s | 1485 tok/s | llama.cpp discussion #10879: Performance of llama.cpp with Vulkan · October 2026 |
| Llama 2 7B Chat | Q4_0 | llama.cpp (ROCm (HIP)) | 50 tok/s | 910 tok/s | llama.cpp discussion #15021: Performance of llama.cpp on AMD ROCm (HIP) · October 2026 |
| gpt-oss-120b | MXFP4 | llama.cpp | 47 tok/s | 1000 tok/s | llama.cpp discussion #16578: Performance of llama.cpp on NVIDIA DGX Spark · November 2025 |
Our take
Should you buy the Ryzen AI Max+ 395 for AI?
The Ryzen AI Max+ 395, also known as Strix Halo, combines 16 Zen 5 cores with a 40-CU Radeon 8060S GPU and up to 128GB of LPDDR5X-8000 on a 256-bit bus for 256 GB/s. With 128GB, Windows can assign up to 96GB to the GPU, and Linux can be configured to give it about 120GB. That is enough for MoE models like gpt-oss-120b and GLM-4.5-Air at good speeds, though dense 70B models generate slowly at this bandwidth. Mini PCs with 128GB cost about $2,000 in 2025 but ranged from about $3,300 to $4,350 in October 2026.
What we like
- Up to 96GB of GPU memory on Windows and about 120GB on Linux
- Small, quiet systems with configurable 45 to 120W chip power
- Strong for MoE models such as gpt-oss-120b
- A full x86 PC that also works as a desktop
Watch out for
- 256 GB/s makes dense 70B models slow
- Slower prompt processing than DGX Spark or NVIDIA GPUs
- Prices have risen sharply
Best for: Running 100B-class MoE models at home · Compact, low-power AI servers · Developers who want an x86 box with lots of GPU memory
Specifications
Ryzen AI Max+ 395 specs that matter for AI
- Memory
- 128GB LPDDR5X-8000 (unified)
- GPU-usable memory
- ~96GB by default
- Memory bandwidth
- 256 GB/s
- Memory bus
- 256-bit
- Architecture
- RDNA 3.5 (Strix Halo)
- Power
- 120W (system)
- Interface
- Integrated (mini PCs and laptops)
- Released
- 2025
- Street price (new)
- ~$3,300 (2026-10)
- Software
- ROCm, Vulkan
Windows allows up to 96GB as dedicated VRAM. On Linux the GPU can use about 120GB after changing a kernel setting.
Sources: ComputingForGeeks Ryzen AI Max+ 395 mini PC price comparison, OpenClawDC Strix Halo mini PC buying guide, Compute Market Strix Halo 128GB prices and benchmarks
Cross-shop
Alternatives to the Ryzen AI Max+ 395
Frequently asked questions
How much VRAM does the Ryzen AI Max+ 395 have?
The AMD Ryzen AI Max+ 395 (128GB) has 128GB of unified memory (about 96GB usable by the GPU by default) with 256 GB/s of memory bandwidth on a 256-bit bus.
What is the largest LLM the Ryzen AI Max+ 395 can run?
Fully in GPU memory, the largest model in our catalog that runs comfortably is Qwen3.5 122B-A10B (122B (10B active)) at Q4_K_M, using about 77GB with 8K context at ~18 tokens/sec.
Can the Ryzen AI Max+ 395 run Llama 3.3 70B?
Yes. At Q8_0 it needs about 81GB with 8K context and runs at ~2.6 tokens/sec.
What is the best LLM to run on the Ryzen AI Max+ 395?
Our top pick is Qwen3.5 122B-A10B at Q4_K_M, which fits in 128GB with room for 8K of context and runs at ~18 tokens/sec. See our full list of the best models for the Ryzen AI Max+ 395 for coding, reasoning, and vision picks.
Is the Ryzen AI Max+ 395 good for local AI in 2026?
The cheapest way to run 100B-class MoE models in one quiet box, though prices have climbed past $3,300.