• Why consider it: Socketed RAM lets you start with 16GB and upgrade later. Reported speeds are 18 to 25 tokens per second on 7B and 8B models.
• Pros: Upgradeable memory, solid price-to-performance, good Ollama and llama.cpp compatibility.
• Cons: Fan noise under sustained load, and 32GB configurations are pricier because of DRAM costs.

Students should also see our best mini PCs for students in 2026.

Third Choice: GMKtec EVO-X2: The Strix Halo Mini PC for Large Models

• Why consider it: This kind of Strix Halo mini PC is among the cheapest ways to hold a 70B model in memory. Community testing reported roughly 5 tokens per second on a dense 70B model and 72 to 75 tokens per second on Mixture of Experts models such as Qwen3-30B-A3B.
• Pros: Huge memory pool, strong MoE speed, 2.5GbE for serving other devices.
• Cons: Higher price, fan noise under AI load, and a Vulkan and ROCm learning curve.

Fourth Choice: MINISFORUM MS-S1 MAX: Workstation-Class Local AI

• Why consider it: It shares the platform that makes 70B-class models practical on a desktop. A third-party tracker put Ryzen AI Max+ 395 systems at 60 to 120W under LLM load, so plan for the power draw.
• Pros: Large shared memory pool, strong GPU for its size, Minisforum’s established U.S. availability.
• Cons: Priced above the EVO-X2 at launch, with the same noise and software-setup trade-offs.

Fifth Choice: Beelink Mini S13: Best for Learning on a Budget

• Why consider it: It runs 8B models at about 6 to 9 tokens per second, which is slow but usable for summaries and a personal coding helper.
• Pros: Very low cost, near-silent in one long stress test, easy to repurpose later.
• Cons: 16GB ceiling, single-channel memory limits bandwidth, not fast enough to feel like a cloud chatbot.

Undecided between a desktop and a notebook? Our mini PCs vs laptops comparison covers the trade-offs, and AMD vs Intel mini PCs explains the platform differences.

Local LLM Hardware Requirements: What Actually Matters

The core local LLM hardware requirements come down to memory. A language model reads its entire weight file from memory for every token it generates. That makes memory capacity the gate (can the model fit at all?) and memory bandwidth the throttle (how fast does it respond?).

The relationship is close to linear. Roughly double the bandwidth and you roughly double your tokens per second. The table below shows the main hardware tiers using figures reported in spring 2026 testing:

Before you install anything, match a model to your memory. Parameters are the “size” of a model, and quantization shrinks the weights so they fit in less memory.

At Q4 (4-bit) quantization, a 13B model occupies roughly 8GB, and a 70B model needs about 40GB. An 8B model such as Llama 3.1 8B is typically around 5GB. These are weights only. The operating system and your KV cache (the memory used for conversation history) need room too.

Three terms will come up constantly:

• Quantization: Lower bit-widths use less memory and run faster, with some quality loss. Q4 is the common balance.
• GGUF: The file format used by llama.cpp and many desktop tools for quantized models.
• Context window: How much text the model can consider at once. Longer context grows the KV cache and uses more RAM.

Mixture of Experts (MoE) models are a useful exception. They activate only a fraction of their parameters per token, so they run far faster than dense models of similar total size. That is why Strix Halo hardware reports such strong MoE speeds.

Prepare Your Mini PC Before Installing Anything

Ollama is the simplest starting point. It handles model downloads, quantization selection, and inference in one command-line tool that works on Linux, macOS, and Windows. Here is how to run a local AI model on a mini PC from a clean install.

Step 1: Install Ollama

Choose your platform:

• Windows and macOS: Download the installer from ollama.com and run it.
• Linux: Open a terminal and run:

curl -fsSL https://ollama.com/install.sh | sh

Confirm the install worked:

ollama --version

Step 2: Download and run your first model

Start with an 8B model, which fits comfortably on 16GB machines:

ollama run llama3.1:8b

Ollama downloads the model on the first run and then opens an interactive prompt. Type a question and press Enter. To exit, type /bye.

Step 3: Manage your models

ollama list        # show downloaded models
ollama pull <name> # download without running
ollama ps          # show loaded models and whether they use GPU or CPU

The ollama ps output is the quickest way to check whether your GPU is doing the work or whether the model has fallen back to the CPU.

Step 4: Test the local API endpoint

Ollama serves an API endpoint on port 11434 by default. Test it:

curl http://localhost:11434/api/generate -d '{"model":"llama3.1:8b","prompt":"Explain unified memory in one sentence.","stream":false}'

This is the same endpoint that chat interfaces, coding assistants, and scripts connect to.

Speed depends on the backend your software uses.

• AMD mini PCs (Windows or Linux): llama.cpp with a Vulkan backend is a common, stable choice across Radeon integrated graphics, from the 780M to the 8060S in Strix Halo systems. ROCm can offer higher theoretical performance for long-context work but needs more setup. Check the release notes for your tool and GPU to confirm current support.
• Apple Silicon: Metal acceleration works out of the box, and MLX is Apple’s native framework for efficient inference. It works with Ollama and LM Studio without driver setup.
• Intel N-series and Core Ultra: Expect CPU-based inference for most tools. Stick with 7B to 8B models and Q4 quantization.

After any change, run ollama ps while a model is loaded and confirm it reports GPU use. If it reports CPU, update your graphics driver and re-test.

Frequently Asked Questions

What is the best mini PC for local AI?

It depends on your model size. For most people the Mac mini M4 Pro is the polished, quiet choice, while a Ryzen 7 8845HS system such as the MINISFORUM UM880 Plus is the value pick. For 70B-class models, a 128GB Strix Halo machine is the practical option.

How much RAM do I need to run a local AI model?

Start with 16GB for 7B to 8B models and 32GB for 13B models. For 70B models, plan on 64GB at minimum and 128GB for comfort. These local LLM hardware requirements also include room for the OS and context cache.

Can I install Ollama on a Windows mini PC?

Yes. Download the Windows installer from ollama.com, run it, then open a terminal and use ollama run llama3.1:8b. You can install Ollama on macOS and Linux as well.

Can I run an LLM offline?

Yes. After the model is downloaded, you can run an LLM offline with no internet connection, which suits private documents and air-gapped setups.

Is a Strix Halo mini PC worth the money?

If you need 70B-class models or heavy agent workflows, a Strix Halo mini PC is among the most affordable ways to get that much memory in a desktop. For 7B to 13B work, a cheaper Ryzen 7 8845HS system or a Mac mini is usually enough.