Bektursun's AI Hardware Store

Open-source models worth running

Sourced from the Arena AI Agent Arena leaderboard, filtered to models with an open license (MIT, Apache-2.0, or a clearly-labeled open variant) and roughly 100 billion parameters or more.

What does "B parameters" mean? A parameter is one number the model learned during training — one knob among billions that shapes how it responds. "753B" means 753 billion of those numbers. More parameters generally means the model can hold more knowledge and nuance, but every one of those numbers has to physically live in GPU memory while the model runs — which is exactly why parameter count is what decides what hardware you need.
Why we use the total parameter count, not just the "active" ones. All four models below are Mixture-of-Experts (MoE) models: for any single word they generate, only a fraction of the parameters ("active parameters") actually do the math. But which fraction is used changes from word to word — so every parameter has to be sitting in GPU memory at all times, ready to be called on. That means the number that determines hardware requirements is always the total parameter count, never the smaller "active" number.
Tencent (Hunyuan)

Hy3 — 295B parameters

Mixture-of-Experts (192 experts, top-8 routing) · 21B active per token

Apache License 2.0

Fully permissive — includes a patent grant, no attribution-display requirement, no revenue threshold. The most open license in this lineup.

Rank 32 of 49 on the Arena AI Agent Arena leaderboard — leaderboard · model card

Minimum GPU memory needed

295B x 1 GB = 295 GB 295 GB x 1.20 = 354 GB minimum required

Recommended minimum build

3x NVIDIA H200 NVL — see what that means

423 GBCombined GPU memory
$96,000Total hardware price
1,800 WCombined power draw
1.50Homes' worth of power

423 GB covers the 354 GB minimum with a little room to spare. Running this for one hour uses 1.8 kWh — about 2.0% of a 90 kWh EV battery.

A cheaper option exists on paper: 4x NVIDIA RTX PRO 6000 Blackwell Workstation Edition would also clear the memory bar for $64,000. We don't lead with it because NVIDIA RTX PRO 6000 Blackwell Workstation Edition is a workstation card with no datacenter-grade interconnect between units — stacking several of them doesn't give you one coherent fast pool of memory the way NVLink-connected datacenter hardware does. See Clusters Explained.

Request a quote for this build

Z.ai (Zhipu AI)

GLM-5.2 — 753B parameters

Mixture-of-Experts (256 experts, 8 routed + 1 shared active per token) · 40B active per token

MIT License

Genuinely MIT per the official Hugging Face model card — unrestricted commercial use, modification, and redistribution.

Rank 14 of 49 on the Arena AI Agent Arena leaderboard — leaderboard · model card

Minimum GPU memory needed

753B x 1 GB = 753 GB 753 GB x 1.20 = 903.6 GB minimum required

Recommended minimum build

7x NVIDIA H200 NVL — see what that means

987 GBCombined GPU memory
$224,000Total hardware price
4,200 WCombined power draw
3.50Homes' worth of power

987 GB covers the 903.6 GB minimum with a little room to spare. Running this for one hour uses 4.2 kWh — about 4.7% of a 90 kWh EV battery.

A cheaper option exists on paper: 10x NVIDIA RTX PRO 6000 Blackwell Workstation Edition would also clear the memory bar for $160,000. We don't lead with it because NVIDIA RTX PRO 6000 Blackwell Workstation Edition is a workstation card with no datacenter-grade interconnect between units — stacking several of them doesn't give you one coherent fast pool of memory the way NVLink-connected datacenter hardware does. See Clusters Explained.

Request a quote for this build

Moonshot AI

Kimi K2.7-Code — 1,000B parameters

Mixture-of-Experts (384 experts, top-8 routing + 1 shared expert) · 32B active per token

Modified MIT License

Functions as plain MIT for almost everyone. The only added condition: a commercial deployment exceeding 100 million monthly active users OR $20M/month in revenue must display "Kimi K2" in its product UI.

Rank 25 of 49 on the Arena AI Agent Arena leaderboard — leaderboard · model card

Minimum GPU memory needed

1000B x 1 GB = 1000 GB 1000 GB x 1.20 = 1200 GB minimum required

Recommended minimum build

9x NVIDIA H200 NVL — see what that means

1,269 GBCombined GPU memory
$288,000Total hardware price
5,400 WCombined power draw
4.50Homes' worth of power

1,269 GB covers the 1,200 GB minimum with a little room to spare. Running this for one hour uses 5.4 kWh — about 6.0% of a 90 kWh EV battery.

A cheaper option exists on paper: 13x NVIDIA RTX PRO 6000 Blackwell Workstation Edition would also clear the memory bar for $208,000. We don't lead with it because NVIDIA RTX PRO 6000 Blackwell Workstation Edition is a workstation card with no datacenter-grade interconnect between units — stacking several of them doesn't give you one coherent fast pool of memory the way NVLink-connected datacenter hardware does. See Clusters Explained.

Request a quote for this build

DeepSeek

DeepSeek-V4-Pro — 1,600B parameters

Mixture-of-Experts · 49B active per token

MIT License

Genuinely MIT per DeepSeek's Hugging Face repo — unrestricted use, modification, and redistribution.

Rank 28 of 49 on the Arena AI Agent Arena leaderboard — leaderboard · model card

Minimum GPU memory needed

1600B x 1 GB = 1600 GB 1600 GB x 1.20 = 1920 GB minimum required

Recommended minimum build

14x NVIDIA H200 NVL — see what that means

1,974 GBCombined GPU memory
$448,000Total hardware price
8,400 WCombined power draw
7.00Homes' worth of power

1,974 GB covers the 1,920 GB minimum with a little room to spare. Running this for one hour uses 8.4 kWh — about 9.3% of a 90 kWh EV battery.

A cheaper option exists on paper: 20x NVIDIA RTX PRO 6000 Blackwell Workstation Edition would also clear the memory bar for $320,000. We don't lead with it because NVIDIA RTX PRO 6000 Blackwell Workstation Edition is a workstation card with no datacenter-grade interconnect between units — stacking several of them doesn't give you one coherent fast pool of memory the way NVLink-connected datacenter hardware does. See Clusters Explained.

Request a quote for this build