What Can You Actually Run on a 64GB or 128GB Local AI Workstation?
A direct comparison of our 64GB ($2,699) and 128GB ($3,899) AI Workstation configs: what model sizes each handles and who each is actually for.
One platform, two memory versions: 64GB at $2,699 or 128GB at $3,899. Same AMD Ryzen AI Max+ 395 CPU/NPU, same Radeon 8060S graphics, same cooling, same ports. Memory is the only thing that changes. Because this platform runs on unified memory, CPU, GPU, and NPU sharing one pool, that one difference matters more than almost anything else on the spec sheet.
The shared foundation
Both configs ship with:
AMD Ryzen AI Max+ 395: up to 50 TOPS on the dedicated NPU, up to 126 TOPS total across CPU, GPU, and NPU
LPDDR5x-8000MT/s unified memory
2TB of PCIe 4.0 storage
Ubuntu Desktop 24.04 or Windows 11 Pro, your choice
LM Studio preinstalled; Ollama and llama.cpp available through our AI Tooling Package
Full SSH and root/administrator access, nothing locked down
Our standard 3-year parts and labor warranty
For the differences between those three tools, see our comparison guide.
64GB: roughly 48GB available for model + context
Practical range: 20B to 35B parameters, think gpt-oss-20b or Qwen’s 32B class, at 4-bit with room to spare for context, or 8-bit if you’d rather trade some headroom for accuracy.
What that’s actually good for:
A private assistant for drafting, summarizing, and answering questions against your own documents (RAG)
AI-assisted development: a real coding assistant running entirely on hardware you control
Light image generation alongside your LLM work
One power user, or light multi-user use on the same box
Start here unless you already know you need 70B or bigger.
128GB: roughly 96GB available for model + context
Practical range: 70B and up, plus large mixture-of-experts models, depending on quantization and context. Dense 70B models and large MoE models both run well here, even a 100B+ MoE model, since only a fraction of its parameters activate per token.
What the extra headroom actually buys you:
Genuinely stronger reasoning and output quality, not just "the same model but bigger"
Comfortable multi-user capacity: several people hitting the same assistant without starving each other for memory
Room to run a larger model and image generation at the same time, no juggling
Real headroom for longer context windows: more of a document, or a longer conversation, held in memory at once
The honest way to decide
If you already know your model class, say you tried a 32B model on a cloud instance and it worked, 64GB is probably enough. If you’re not sure, or you know your needs will grow (more users, bigger documents, harder reasoning tasks), 128GB buys you real headroom so you’re not outgrowing the machine in a year.
Neither config is "the real one," with the other as some lesser version; they’re built for different jobs. If you want to work out your own VRAM number first, the VRAM sizing guide walks through the general calculation.