WizardLM-2 8x22B: Specs and Hardware Needs for the MoE Model

Looking for an open-weight model that can hold up on hard math and logic? WizardLM-2 8x22B is the tenth pick in this uncensored, open-weight series, and it is the heaviest one yet. It comes from the Microsoft AI research team and stacks eight 22B sub-experts into a single mixture-of-experts model.

That gets you scientific deduction, math, and tangled logic close to the frontier proprietary models, with a 64k token context window. The catch is the hardware bill, and it is a big one.

What Happened

• WizardLM-2 8x22B is a mixture-of-experts build with eight 22B experts, made by the Microsoft AI research team.
• VRAM: 2x 80GB for FP8, or 4x 24GB for Q4, or a Mac Studio with 128GB+ unified memory.
• System RAM: 256GB.
• Storage: about 75GB for the GGUF Q4_K_M file, up to 280GB for the full weights.
• Context window: 64k tokens.

Why It Matters

The payoff is reasoning quality: science, math, and complicated logic that this model pushes near what the closed frontier models do. What it costs is a workstation. Two 80GB cards or a 256GB-RAM machine is not a laptop setup, and 280GB of disk is a real amount of space to give up, so check your build before you commit. If your hardware fits, you get a 64k-context model you run yourself.