AI in a Box: Massive gaming capability, heavy-duty AI compute, and standard PC parts, all in 4.5L.
The Framework Desktop with the AMD Ryzen AI Max+ 395 processor and 128GB of soldered LPDDR5x-8000 memory is the ultimate AI workstation in a mini PC form factor. Released in 2025, it features the powerful Radeon 8060S integrated GPU with up to 96GB of dedicated VRAM (in Windows), making it capable of running very large language models like OpenAI's gpt-oss-120b at 38 tokens/second.
Starting at $1,269 (for the base 32GB model). DIY build-yourself PC with moderate difficulty and approximately 10 minutes setup time. Supports Windows 11 or any Linux distribution.
| Spec | Details |
|---|---|
| Processor | AMD Ryzen AI Max+ 395 (soldered) |
| Base Clock | 3.0GHz |
| Max Boost | Up to 5.1GHz |
| Cores / Threads | 16-core / 32-thread |
| L3 Cache | 64MB |
| Processor Power | 120W sustained, 140W boost |
| Memory | 128GB LPDDR5x-8000 (soldered) |
| Memory Bus | 256-bit at 8000 MT/s |
| GPU | AMD Radeon 8060S |
| GPU Clock | Up to 2.9GHz |
| Compute Units | 40 CUs |
| MALL Cache | 32MB |
| NPU | 32 Tiles, up to 50 TOPS |
| Storage | 2x Samsung 990 EVO Plus SSD 4TB |
| Form Factor | Mini-ITX, 4.5L volume |
| Weight | 3.1 kg |
| Case Dimensions | 96.8 x 205.5 x 226.1 mm (H x W x D) |
| Power Supply | 400W FlexATX, ATX 3.0, 80 Plus Gold (110V) / Silver (230V) |
| PSU Fan | Delta AFB0412SHBYQB 40x40mm (0-RPM mode) |
| OS | archlinux |
| Release Year | 2025 |
| Fan | Speed | Noise | Airflow | Connector |
|---|---|---|---|---|
| Noctua NF-A12x25 HS-PWM | 2400 RPM | 28.8 dBA | 117.6 m³/h (69.25 CFM) | 4-pin PWM |
With up to 96GB of memory accessible by the Radeon 8060S GPU (even more on Linux), very large language models like OpenAI's gpt-oss-120b can run real-time.
| Model | Quantization | Speed |
|---|---|---|
| OpenAI gpt-oss-20b | MXFP4 | 58 tok/s |
| OpenAI gpt-oss-120b | MXFP4 | 38 tok/s |
| Configuration | Total Memory | Max Dedicated VRAM (Windows) |
|---|---|---|
| Max 395+ (128GB) | 128GB | 96GB |
* In Linux you can override the VRAM setting to go higher.
LM Studio, Ollama, llama.cpp, and other open source libraries work out of the box on Windows and Linux:
Using kyuz0's AMD Strix Halo toolboxes for optimized llama.cpp inference on this hardware.
VS Code with AI coding assistants:
| Model | Status |
|---|---|
| Qwen3.5-122B-A10B | Old champion |
| Qwen3.6-35B-A3B | Currently testing |
| Gemma-4-26B-A4B-IT | — |