Aydar's AI Hardware Co.

Supported Open-Source AI Models

These are real, currently available open-source models with permissive MIT or Apache-2.0 licenses. Each one is large enough that running it requires serious GPU memory — the calculation below shows exactly how much, using the same formula for every model on this site.

MIT License

GLM 5.2 (Max)

Developer: Z.ai (Zhipu AI)

A frontier-scale mixture-of-experts model from the GLM family, ranked on Agent Arena's open-source tier — only 40B of its 753B parameters activate per token, but all 753B still have to sit in GPU memory at once.

753 billion total parameters (40B active per token, mixture-of-experts)

Parameter count is the size of the model file itself — more parameters generally means smarter answers, but also a much bigger chunk of GPU memory to hold it.

Required GPU memory

Base memory: 753 × 1 GB = 753.0 GB 20% working room: 753 × 0.20 = 150.6 GB Minimum total memory: 753.0 + 150.6 = 903.6 GB
The extra 20% is working room — space the model needs while it's actively answering a question, on top of just storing its own weights.

Minimum recommended hardware setup

NVIDIA GB200 NVL72 Rack — 13.40 TB combined GPU memory (903.6 GB needed, so there's a small safety margin)

Total hardware price: $3,000,000

Total hardware power draw: 130,000 W — about 108.33 average homes worth of continuous power.

Source · verified 2026-07-20

MIT License

DeepSeek V4 Pro

Developer: DeepSeek

A 1.6-trillion-parameter mixture-of-experts model — one of the largest open-weight models tracked on Agent Arena, activating 49B parameters per token.

1600 billion total parameters (49B active per token, mixture-of-experts)

Parameter count is the size of the model file itself — more parameters generally means smarter answers, but also a much bigger chunk of GPU memory to hold it.

Required GPU memory

Base memory: 1600 × 1 GB = 1600.0 GB 20% working room: 1600 × 0.20 = 320.0 GB Minimum total memory: 1600.0 + 320.0 = 1920.0 GB
The extra 20% is working room — space the model needs while it's actively answering a question, on top of just storing its own weights.

Minimum recommended hardware setup

NVIDIA GB200 NVL72 Rack — 13.40 TB combined GPU memory (1920.0 GB needed, so there's a small safety margin)

Total hardware price: $3,000,000

Total hardware power draw: 130,000 W — about 108.33 average homes worth of continuous power.

Source · verified 2026-07-20

MIT License

MiMo V2.5 Pro

Developer: Xiaomi

A trillion-parameter open mixture-of-experts model built to match frontier coding and agent performance, activating 42B parameters per token.

1020 billion total parameters (42B active per token, mixture-of-experts)

Parameter count is the size of the model file itself — more parameters generally means smarter answers, but also a much bigger chunk of GPU memory to hold it.

Required GPU memory

Base memory: 1020 × 1 GB = 1020.0 GB 20% working room: 1020 × 0.20 = 204.0 GB Minimum total memory: 1020.0 + 204.0 = 1224.0 GB
The extra 20% is working room — space the model needs while it's actively answering a question, on top of just storing its own weights.

Minimum recommended hardware setup

NVIDIA GB200 NVL72 Rack — 13.40 TB combined GPU memory (1224.0 GB needed, so there's a small safety margin)

Total hardware price: $3,000,000

Total hardware power draw: 130,000 W — about 108.33 average homes worth of continuous power.

Source · verified 2026-07-20