Listen to this article:
AMD’s Strix Halo mini PCs pack 128GB of unified memory that both the CPU and GPU can access, enough to run 70-billion-parameter language models locally, all in a box that fits on your desk and draws less power than a gaming PC. The LocalLLaMA community and independent benchmarkers have been testing them for a year.
Memory capacity decides which models fit, and memory bandwidth decides how fast they run. These chips share up to 128GB of unified memory between the CPU and GPU, so large models fit without splitting them across cards. When you run a language model on a typical RTX 4090, you’re limited to 24GB of VRAM, and if your model doesn’t fit, you’re either quantizing aggressively or suffering through CPU offloading that tanks performance. Strix Halo’s Ryzen AI Max+ 395 can give up to 96GB of its 128GB system memory to the GPU through AMD’s standard allocation setting, and Linux setups can assign more as GTT memory, which removes that constraint.
Bandwidth Sets the Token Rate
The flagship Ryzen AI Max+ 395 has 16 Zen 5 CPU cores, 40 RDNA 3.5 compute units in the integrated Radeon 8060S GPU, and an XDNA 2 NPU rated at 50 TOPS for dedicated AI operations. These processors support LPDDR5X-8000 memory with around 256 GB/s theoretical bandwidth, though real-world testing shows about 212-215 GB/s in practice.
Token generation in language models is almost entirely memory-bandwidth-bound: the GPU needs to read the model weights from memory for each token it generates. With 215 GB/s of bandwidth feeding 40 compute units, you get roughly 5 tokens per second on a 70B parameter model, scaling up or down predictably based on model size and quantization level.
For context, that 5 tok/s on a 70B model is slow enough that you’ll notice delays in conversation, but fast enough to be usable for interactive chat. The community consensus is that 3-5 tok/s is the minimum for tolerable conversation. Drop down to 30B or 13B models, and the experience becomes noticeably snappier.
Mixture of Experts models run far faster, because only a subset of their parameters activates for each token. In Level1Techs forum testing, a Qwen MoE model with 30B total and 3B active parameters generates about 70 tokens per second, while a dense 70B model manages 4.5.
The Software Learning Curve
AMD’s ROCm software stack needs more setup than CUDA. The performance you get from a Strix Halo system depends heavily on which software backend you use and on your context length.
For general-purpose LLM chat, the Vulkan backend with the RADV driver delivers the best experience. It provides consistent performance across different model types. The ROCm/HIP backend offers higher theoretical performance in some scenarios, but requires more setup and can be finicky with certain configurations.
Vulkan performance degrades significantly as context grows beyond 4,000 tokens, while a properly configured ROCm setup with Flash Attention maintains consistent performance even at 8,000+ tokens. If you’re building RAG pipelines or doing document analysis that requires long context, ROCm is worth the setup pain.
A common setup is Fedora 42 or a recent Ubuntu distribution with the latest Mesa drivers, running llama.cpp with Vulkan for general use and switching to ROCm for long-context work. This is more complexity than you’d deal with on an NVIDIA system, where CUDA just works, but it’s manageable for anyone willing to follow documentation and ask an AI assistant for help with configuration.
What Strix Halo Handles Well
A local coding assistant works well enough for typical code-completion and explanation tasks. A mixture-of-experts coding model with about 3B active parameters runs at the MoE speeds above, while a dense 70B model answers at about 5 tok/s, noticeably slower than a cloud API.
For those building RAG applications, the combination of large memory and reasonable inference speed makes Strix Halo practical. You can load a substantial model, keep a vector database in memory, and process queries without the orchestration complexity of multi-GPU setups or the latency penalties of cloud APIs. The Medium piece from Orami reports that for developers who need local LLM capability for privacy, cost control, or offline access, these systems finally deliver workstation-class capability.
LoRA and QLoRA fine-tuning on 20-30B models fits in 128GB of unified memory, letting you customize models for specific domains or tasks without cloud compute costs. While it’s not as fast as training on an H100, it’s fast enough for experimentation and small-scale customization.
Throughput is the limitation. If you’re serving multiple concurrent users, processing large batches of documents, or running inference as part of a production pipeline, Strix Halo’s roughly 70 tok/s on a 30B MoE model and 5 tok/s on a 70B dense model won’t compete with dedicated AI accelerators or cloud instances.
The NVIDIA DGX Spark Comparison
NVIDIA raised the DGX Spark Founders Edition from $3,999 to $4,699 in February 2026, citing memory supply. As of September 2026, 128GB Strix Halo mini PCs list for about $3,100 (the BOSGAME M5) to $4,350 (the Beelink GTR9 Pro), and AMD’s own Ryzen AI Halo developer box sells for $3,999. So the Spark costs about $350 to $1,600 more. Both systems target local AI development with 128GB of unified memory. The Spark buys faster prompt processing, more polished software, and CUDA; Strix Halo buys x86 compatibility at a lower price.
The DGX Spark uses NVIDIA’s GB10 Grace Blackwell superchip with Arm cores, and it reads prompts much faster. In llama.cpp’s own DGX Spark benchmarks, gpt-oss-120b processes a 2,048-token prompt at about 2,440 tokens per second. Donato Capitella’s Strix Halo results show about 720 tokens per second for the same model on a shorter 512-token prompt, so the Spark is more than three times faster at this step. On long documents or large codebases, the Spark starts answering sooner.
Token generation speeds are close. The same two sources show about 59 tokens per second on the Spark and about 57 on Strix Halo for gpt-oss-120b, so interactive chat feels nearly identical on both.
The software ecosystem tilts heavily toward NVIDIA. DGX Spark comes with a curated AI development environment that just works, while Strix Halo requires more manual configuration and occasional troubleshooting. If your time is valuable and you want the smoothest possible experience, the DGX Spark’s premium might be worth it.
Pick the DGX Spark if you need CUDA or fast prompt processing on long documents. For interactive chat and x86 software, Strix Halo matches its generation speed for less money.
Available Hardware Options
The boxes below differ mainly in networking, expansion, and support. If you’re already familiar with the MINISFORUM MS-01 or similar high-performance mini PCs, Strix Halo represents the next step up for AI-focused workloads.
MINISFORUM MS-S1 MAX
- +USB4 V2 80Gbps connectivity
- +Dual 10GbE networking
- +Built-in 320W PSU
- +PCIe x16 expansion slot
- -Premium pricing
- -RAM soldered (non-upgradeable)
- -Larger footprint than competitors
GMKtec EVO-X3
- +128GB LPDDR5X-8000 standard
- +OCuLink PCIe Gen4 x4 for an external GPU
- +USB4
- +WiFi 7
- -RAM soldered (non-upgradeable)
- -Single 2.5GbE port
- -Power limits locked in firmware
HP Z2 Mini G1a
- +HP enterprise support and warranty
- +Thunderbolt 4 ports
- +Flex IO customization
- +ISV-certified for professional apps
- -This PRO 385 build has only 32GB
- -Mini DisplayPort only (no full-size HDMI)
- -2.5GbE instead of 10GbE
GMKtec EVO-X2
- +Multiple RAM configurations (64/96/128GB)
- +Quick-release SSD design
- +Quad 8K display support
- -Dual 2.5GbE (not 10GbE)
- -RAM soldered
- -Inconsistent availability
Two more boxes build on the same chip. The Beelink GTR9 Pro pairs it with dual Intel 10GbE, WiFi 7, and DisplayPort 2.1 for 8K output, and the NIMO Ryzen AI Max+ 395 fits it into a 4L chassis with configurable power modes up to 120W.
The Framework Desktop offers 64GB and 128GB configurations. Framework’s emphasis on repairability and modularity appeals to those who want maximum flexibility, and the 4.5-liter chassis is notably compact. Framework actively contributes to driver development. For a broader look at available hardware, our mini PC comparison chart lets you filter models by the specs that matter for AI workloads.
The Total Cost Picture
At an example rate of $0.015 per thousand input tokens and $0.03 per thousand output tokens, a developer running 10,000 tokens per hour of inference spends at most $2.40 over an eight-hour day (if every token bills at the output rate), or roughly $600 over a year of workdays.
At that rate, a 128GB Strix Halo box listing for $3,100 to $4,350 takes at least five to seven years to pay for itself in API savings. It pays back sooner only for usage several times heavier. The electricity cost is small next to the hardware, with the system drawing 120-140W under inference load. For most buyers, the case rests on privacy and offline use, not on savings.
The calculation is worse for lighter users. If you’re only occasionally running local models or primarily using cloud APIs for production work, the upfront cost is harder to justify, especially with DRAM prices surging across the industry. At lighter use the payback period grows in proportion, so cloud APIs cost less, even before counting the time spent on local setup and maintenance.
Strix Halo Suits Developers Wanting Local LLMs
You should be comfortable with Linux (distributions like Omarchy have made the experience more accessible), willing to troubleshoot occasional driver or backend issues, and patient with the 5 tok/s experience on larger models.
The worst-case Strix Halo buyer is someone expecting a plug-and-play ChatGPT replacement. The software ecosystem, while improving rapidly, still requires more hands-on attention than NVIDIA alternatives. If you’re not willing to learn about backends, quantization levels, and context management, you’ll find the experience frustrating regardless of how capable the hardware is. Developers looking for AI capabilities with less complexity are also exploring dedicated AI agent setups that use cloud APIs for inference while local hardware handles orchestration.
For those primarily interested in smaller models, 7B to 13B parameters, a more modest investment might serve you better. An N100 or N150 mini PC can run the 7B end of that range for a fraction of the cost, though its single memory channel leaves 13B models near the 3-5 tok/s floor, and the Mac Mini M4 offers another compelling option for those who prefer Apple’s ecosystem. Reserve Strix Halo’s substantial investment for users who need to run 30B, 70B, or larger models locally. For a mid-range option that bridges the gap, the MINISFORUM AI X1 Pro offers strong AI performance at a lower price point than full Strix Halo systems.
AMD’s Gorgon Halo refresh (Ryzen AI Max 400) raises the unified-memory ceiling to 192GB with slightly faster LPDDR5X-8533 on the same 256-bit bus, and the first systems are now on sale: MINISFORUM’s MS-S1 MAX-P495 lists at $7,399 with 192GB. At that price it makes sense only for models that do not fit in 128GB. For everything else, a 128GB Strix Halo box is the better buy.
