Open-source models we size hardware for
All three are MIT-licensed open-weights models. A model's size is its parameter count — think of each parameter as one small number the model learned during training. More parameters generally means a more capable model, but every parameter has to live in GPU memory while the model runs, which is why parameter count is the number that decides what hardware you need.
GLM-5.2
A frontier-scale open-weights model. At 753 billion parameters it sits near the top of the open leaderboard, which is exactly why it needs the memory math below taken seriously.
753 billion parameters
Minimum build: 1× NVIDIA H200 SXM Server (8-GPU) (1128 GB total, $380,000)
Full memory mathGLM-5.1
The prior release in the same family as GLM-5.2, essentially identical in size at 754 billion parameters — a reminder that "which version" matters less than "how many parameters" when you're sizing hardware.
754 billion parameters
Minimum build: 1× NVIDIA H200 SXM Server (8-GPU) (1128 GB total, $380,000)
Full memory mathDeepSeek V4 Flash
A lighter, faster sibling in the DeepSeek V4 family. At 284 billion parameters it needs well under half the memory of GLM-5.x, which is why it's the model to reach for when a single 8-GPU server should be enough.
284 billion parameters
Minimum build: 1× NVIDIA H100 SXM Server (8-GPU) (640 GB total, $300,000)
Full memory mathHow many GPUs does it fill?
One tile = one H200 GPU (141 GB). Pick a model to see how many it takes.
Compare two models
Bars scaled to one 8-GPU H200 server (1128 GB total).