The AMD Ryzen AI Max+ 395, known by its codename Strix Halo, is the cheapest way in 2026 to get around 96 GB of memory that a GPU can use.
Mini PCs built on it come with 128 GB of unified memory, run gpt-oss-120b at over 30 tokens per second, and cost about half as much as an NVIDIA DGX Spark.
The catch is memory bandwidth. At about 256 GB/s, big dense models like Llama 70B crawl at 5 to 7 tokens per second. This guide covers the specs, every 128 GB mini PC with current prices, real benchmarks, what runs well, how to set it up and who should buy one.
- What it is: a 16-core Zen 5 chip with a 40-CU Radeon 8060S GPU and up to 128 GB of shared LPDDR5X memory.
- Usable GPU memory: up to 96 GB on Windows, roughly 96-120 GB on Linux.
- Speed: gpt-oss-120b about 31-34 t/s, Qwen3 30B-A3B about 72-75 t/s, dense 70B only about 5-7 t/s.
- Price: 128 GB mini PCs from about $3,449 (Framework, DIY) and $3,599 (Bosgame M5), versus $6,950 for a 128 GB DGX Spark.
- Best for: big mixture-of-experts (MoE) models on a budget. Not for fast dense 70B models or CUDA-only tools.
Ryzen AI Max+ 395 Specs for AI
| Spec | Ryzen AI Max+ 395 | Why it matters |
|---|---|---|
| CPU | 16 Zen 5 cores | Fast for CPU-side work and tools |
| GPU | Radeon 8060S, 40 compute units | Runs models through Vulkan or ROCm |
| Memory | Up to 128 GB LPDDR5X-8000, 256-bit | One shared pool for CPU and GPU |
| Bandwidth | About 256 GB/s on paper, about 215 GB/s measured | The main limit on text generation speed |
| GPU memory | Up to 96 GB on Windows; about 96-120 GB on Linux | Fits 100B+ MoE models |
| NPU | 50 TOPS (XDNA 2) | Copilot+ features; little use for LLMs today |
| Power | About 120-140 W sustained in most mini PCs | Quiet and cheap to run compared with a big GPU |
There is also a 192 GB version, the Ryzen AI Max+ PRO 495, with about 273 GB/s and up to 160 GB for the GPU. AMD sells it in its own Ryzen AI Halo developer box for $7,299.99 at Micro Center.
Every 128 GB Strix Halo Mini PC and Its Price
These prices were checked by Computing for Geeks on 22 October 2026. Strix Halo machines sell out often, so check current stock before you decide.
| Machine | Price (128 GB, 2 TB) | Networking | Notes |
|---|---|---|---|
| Framework Desktop | $3,449 (DIY, no SSD) | 5GbE | Mini-ITX, repairable, PCIe slot; often out of stock |
| Bosgame M5 | $3,599 | 2.5GbE | Cheapest 128 GB box in stock; external 240 W brick |
| GMKtec EVO-X2 | $3,649.99 (64 GB/1 TB: $2,199.99) | 2.5GbE | Loud under long loads |
| GMKtec EVO-X3 | $3,799.98 | Not stated | Adds an OCuLink port for an external GPU |
| Minisforum MS-S1 Max | $3,879 | Dual 10GbE | Two 80 Gbps USB4 v2 ports, PCIe x16 slot (x4 electrical), rack-mountable |
| Beelink GTR9 Pro | $4,349 | Dual 10GbE | Internal PSU, vapour chamber cooling |
| AMD Ryzen AI Halo | $3,999.99 | Not stated | AMD’s own developer box |
If you plan to link two boxes together for bigger models, pick one with 10GbE or USB4 v2, such as the Minisforum MS-S1 Max or Beelink GTR9 Pro.
These prices are far above early 2025 levels because LPDDR5X memory costs have risen sharply. My guide to why RAM is so expensive in 2026 explains what’s driving it.
LLM Benchmarks: Real Tokens per Second
I haven’t tested a Strix Halo machine myself. These numbers come from Level1Techs, ServeTheHome, Hardware Corner, StorageReview and HotHardware, as compiled by Presenc AI in October 2026.
| Model | Quant | Prompt (t/s) | Generation (t/s) | Source |
|---|---|---|---|---|
| Llama 2 7B | Q4_0, Vulkan | 998 | 45.8 | Level1Techs |
| Qwen3 30B-A3B (MoE) | Vulkan | 605 | 72.0 | Level1Techs |
| gpt-oss-120b (MoE) | MXFP4, llama.cpp | 340 | 34.1 | Hardware Corner |
| gpt-oss-120b (MoE) | LM Studio | n/a | 31.4 | ServeTheHome |
| Llama 3.1 70B (dense) | 4-bit | n/a | 6 to 7 | HotHardware |
| Dense 70B | 4-bit | n/a | About 5 | Level1Techs, ServeTheHome |
The pattern is clear. MoE models, which only read a small part of their weights for each token, run fast. Dense models read everything and hit the bandwidth wall.
Results change a lot with software. Level1Techs saw a gain of nearly 50% in one prompt test between May and August 2025 from driver and llama.cpp updates alone, and Vulkan and ROCm builds give different numbers.
Serving many users at once
StorageReview tested batched serving with vLLM. On gpt-oss-20b with 64 requests at once, Strix Halo managed 617 t/s against 1,917 t/s on a DGX Spark. For a family or small team that is fine, but the Spark is much stronger for heavy serving.
Which Models Run Well (and Which Don’t)
| Model type | Example | Experience on Strix Halo |
|---|---|---|
| Large MoE, about 100-120B | gpt-oss-120b | Excellent: about 30+ t/s, fits with long context |
| Medium MoE, about 30B | Qwen3 30B-A3B, Qwen3 Coder 30B | Very fast: about 70+ t/s |
| Small dense, 7-14B | Llama 3.1 8B, Qwen3 14B | Good, but a mid-range GPU is faster |
| Dense 27-32B | Qwen3.8-27B | Usable but slow; a 24 GB GPU is better |
| Dense 70B | Llama 3.1 70B | Fits, but only about 5-7 t/s |
| Very large MoE, 200B+ | Bigger open-weight MoE models | Only at low-bit quants or across two linked boxes |
For picking a quant that fits, my how much RAM for local AI guide has the memory needed for each model size.
Strix Halo vs DGX Spark vs Mac vs RTX GPUs
| Option | Memory for models | Bandwidth | gpt-oss-120b | Rough price, Nov 2026 |
|---|---|---|---|---|
| Ryzen AI Max+ 395 mini PC | Up to 96-120 GB | About 256 GB/s | About 31-34 t/s | From about $3,450-$3,600 |
| NVIDIA DGX Spark (128 GB) | 128 GB | About 273 GB/s | Similar single-user speed, far faster batched | $6,950 |
| Used RTX 3090 24 GB | 24 GB | 936 GB/s | Needs heavy offload | About $1,450 for the card |
| Mac with 128 GB unified memory | Most of 128 GB | Higher on Max chips | Fast | Usually more than Strix Halo |
The DGX Spark wins if you need CUDA, fine-tuning tools or serving many users. A Mac wins on polish and battery life in a laptop. Strix Halo wins on price per gigabyte of model memory and runs normal Windows or Linux.
For full breakdowns, see my DGX Spark guide, Mac vs PC for local AI and used RTX 3090 guide.
How to Set Up a Strix Halo PC for Local AI
On Windows
- Update the BIOS and AMD Adrenalin driver from the mini PC maker and AMD’s site.
- Give the GPU more memory. In AMD Software: Adrenalin Edition, open the performance tuning settings and raise Variable Graphics Memory (some BIOS menus call it UMA frame buffer). Up to 96 GB is possible on a 128 GB machine.
- Install LM Studio and choose the Vulkan or ROCm runtime in its settings. It is the easiest way to start.
- Load a model that fits, such as gpt-oss-120b in MXFP4 or Qwen3 30B-A3B, and offload all layers to the GPU.
- Check the speed and try the other runtime, since Vulkan and ROCm can differ by model.
On Linux
- Use a recent kernel and Mesa for the best Vulkan support.
- llama.cpp with the Vulkan backend is the most common choice among owners; ROCm builds also work on recent ROCm releases.
- Owners raise the GPU’s share of memory through kernel settings for the graphics translation table (GTT), which is how they reach about 96-120 GB.
- Ollama runs, but check that it actually uses the GPU with
ollama ps; if it falls back to the CPU, use LM Studio or llama.cpp directly.
My AMD GPU for AI guide covers ROCm, Ollama and ComfyUI on AMD hardware in more depth.
Leave at least 16-24 GB for the system. If you give the GPU everything, Windows or your browser can start swapping and every model slows down.
Image Generation and Other Workloads
Strix Halo can run ComfyUI through ROCm, and the huge memory means big image and video models fit without offloading. But the 40-CU iGPU is much slower at image generation than a mid-range RTX card, and many ComfyUI custom nodes assume NVIDIA CUDA.
If image and video work is your main goal, an RTX GPU is the better buy. See my best GPU for local AI guide for options by budget.
Should You Buy One?
| You want to… | Verdict |
|---|---|
| Run gpt-oss-120b or other large MoE models at home | Yes, the best value option in 2026 |
| Run a private coding assistant with a 30B MoE model | Yes, very fast |
| Run dense 70B models quickly | No, consider 2-3 GPUs instead |
| Fine-tune models or use CUDA-only tools | No, look at DGX Spark or RTX cards |
| Do mostly image and video generation | No, an RTX GPU is faster |
| Serve many users at once | Only lightly; DGX Spark is far stronger |
If you’d rather have a laptop, a few 2026 laptops also use this chip. My best laptop for local AI guide compares them with MacBooks and RTX laptops.
Troubleshooting
The GPU only shows a small amount of memory
Raise Variable Graphics Memory in AMD Adrenalin, or the UMA frame buffer size in the BIOS, then reboot. On Linux, raise the GTT limit.
LM Studio or Ollama runs on the CPU
Update the AMD driver, select the Vulkan or ROCm runtime, and offload all layers to the GPU. Check usage in Task Manager on Windows or with ollama ps.
Generation is slow on a 70B model
That’s the bandwidth limit, not a fault. Switch to an MoE model like gpt-oss-120b, which runs about five times faster.
The mini PC is loud or throttles
Use the maker’s balanced or quiet power mode. Some boxes, such as the GMKtec EVO-X2, are loud at full power, and the speed drop in quieter modes is often small.
ComfyUI nodes fail with CUDA errors
Some custom nodes only support NVIDIA. Look for ROCm-compatible versions or use core nodes instead.
Frequently Asked Questions
Is the Ryzen AI Max+ 395 good for local LLMs?
Yes, for large mixture-of-experts models. It runs gpt-oss-120b at about 31-34 tokens per second thanks to up to 96 GB of GPU memory, but dense 70B models only reach about 5-7 tokens per second.
How much memory can the GPU use on Strix Halo?
Up to 96 GB on Windows on a 128 GB machine. On Linux owners report about 96-120 GB by raising the GTT limit.
How much does a 128 GB Strix Halo mini PC cost?
In late October 2026 prices started at about $3,449 for the DIY Framework Desktop and $3,599 for the Bosgame M5, with most models between $3,600 and $4,350.
Strix Halo or DGX Spark?
Strix Halo costs about half as much and gives similar single-user speed on gpt-oss-120b. The DGX Spark is better for CUDA tools, fine-tuning and serving many users at once.
Can Strix Halo run Llama 70B?
It fits, but generation is only about 5-7 tokens per second because of the 256 GB/s memory bandwidth.
Does Ollama work on the Ryzen AI Max+ 395?
Ollama runs, but check that it uses the GPU. Many owners get better results with LM Studio or llama.cpp using the Vulkan or ROCm backend.
Is the Ryzen AI Max+ 395 good for Stable Diffusion and ComfyUI?
It works and fits big models, but it is much slower than a mid-range RTX GPU for image generation, and some ComfyUI nodes need CUDA.
Final Thoughts
The Ryzen AI Max+ 395 is the most affordable way to run 100B-class MoE models at home in 2026. If gpt-oss-120b or a fast 30B coding model is what you want, a 128 GB mini PC at around $3,600 is hard to beat.
Just go in with the right expectations: it is a capacity machine, not a speed machine. For dense models, image generation or CUDA work, a GPU-based PC is still the better choice.
Sources
- Computing for Geeks: Ryzen AI Max+ 395 price, cheapest 128 GB mini PC (November 2026)
- Presenc AI: AMD Strix Halo local LLM benchmarks (October 2026)
- PromptQuorum: AMD Ryzen AI Halo specs, price and verdict
- RunAIHome: Ryzen AI Max+ 395 for local LLMs in 2026
- LlamaBuilds: Ryzen AI Max+ 395 specs, speed and what it runs
