FreeLLMAI PCs New
Computers built for AI share one big pool of fast memory between the processor and the graphics chip, so they can hold models far larger than a graphics card. They are quiet and small, but slower than a big graphics card on models that fit in it. Estimates from our simulator.
How fast they write
Tokens per second, estimated. Blue: AI PCs with unified memory. Orange: a desktop PC with a graphics card, for comparison.
What this shows · How to read it
What it shows. Estimated speed on Qwen3.8-27B, a dense model with 27 B parameters, at 4-bit with 8K context. The desktop PC has an RTX 5090 and 64 GB of RAM.
How to read it. Longer bar, faster. Memory bandwidth decides the speed: every word reads the whole model from memory. AI PCs win on size, graphics cards win on speed when the model fits.
Apple Mac Studio with M3 Ultra
macOS: LM Studio, Ollama, llama.cpp and MLX all run on the Apple GPU. On Qwen3.8-27B (a dense 27 B model) it writes about 33 tokens per second. Every model it runs →
- Mac Studio M3 UltraDesktopAmazon ↗
Apple M4 Max: Mac Studio and MacBook Pro
macOS: LM Studio, Ollama, llama.cpp and MLX all run on the Apple GPU. On Qwen3.8-27B (a dense 27 B model) it writes about 23 tokens per second. Every model it runs →
NVIDIA DGX Spark and other GB10 machines
Linux with the NVIDIA CUDA stack: Ollama, llama.cpp, vLLM and most AI tools work as on an NVIDIA graphics card. On Qwen3.8-27B (a dense 27 B model) it writes about 12 tokens per second. Every model it runs →
AMD Ryzen AI Max+ 395 mini PCs and laptops
Windows or Linux: LM Studio, Ollama and llama.cpp run on the built-in Radeon GPU (Vulkan or ROCm). On Qwen3.8-27B (a dense 27 B model) it writes about 11 tokens per second. Every model it runs →
Apple Mac mini with M4 Pro
macOS: LM Studio, Ollama, llama.cpp and MLX all run on the Apple GPU. On Qwen3.8-27B (a dense 27 B model) it writes about 12 tokens per second. Every model it runs →
- Mac mini M4 ProMini PCAmazon ↗
For comparison: a desktop PC with an RTX 5090
32 GB on the graphics card, 1792 GB/s, plus 64 GB of system RAM. Best free model at reading speed: Gemma 4 31B
Every model the RTX 5090 runs → · All graphics cards · How much RAM
How to choose
Memory decides what fits
A model at 4-bit needs about 0.6 GB per billion parameters, plus the context. With unified memory, about three quarters of it can go to the model: 128 GB holds models of about 150 billion parameters, 512 GB even bigger ones.
Bandwidth decides the speed
Every word the model writes reads all its active weights. More GB/s, faster answers. Mixture-of-experts models read only a part of their weights, so they run fast even on big AI PCs.
Apple, NVIDIA or AMD
Macs are silent and simple to set up. GB10 machines like the DGX Spark run the NVIDIA software most AI tools are built for. Ryzen AI Max+ 395 mini PCs are the cheapest way to 128 GB and run Windows or Linux.
Or a graphics card?
If the models you want fit in 24 to 32 GB, a desktop with a big graphics card is faster. Choose an AI PC when you want models that need 64 GB or more. Compare in the Run at home simulator.
Affiliate links: as an Amazon Associate, FreeLLMArena earns from qualifying purchases made through the “Amazon ↗” links. The links do not change the order, the numbers or the estimates on this page, and the price is the same for you. Speeds are estimates from our simulator, calibrated on real llama.cpp runs.
Share this page:XLinkedInRedditWhatsApp
Spotted a wrong number or a broken link? Report a mistake