The Reckoning
Three great open-source models from the world's agent leaderboard (arena.ai/leaderboard/agent), filtered to Open Source — all under the MIT license, all well past 100 billion parameters. For each, the arithmetic that determines the smallest honest engine able to run it, shown in full.
DeepSeek V4 Flash
The smallest of our three flagship models — MoE keeps inference cheap (only 13B parameters active per token) but every expert still has to sit in GPU memory.
What it means: parameter count is the model's size. “284B” means 284 billion weights. Even for a Mixture-of-Experts model where only 13B are active per token, every expert must still reside in memory at once — one cannot decline to load the experts not needed this particular token.
564 GB ≥ 340.8 GB required, making this the smallest true configuration in our house that fits the model. Real cluster: these 4 GPUs share one NVLink memory domain, so their memory truly adds up to 564 GB.
GLM-5.2 (Max)
A much larger MoE model — needs the full 8-GPU NVLink node to have any room to spare.
What it means: parameter count is the model's size. “744B” means 744 billion weights. Even for a Mixture-of-Experts model where only 40B are active per token, every expert must still reside in memory at once — one cannot decline to load the experts not needed this particular token.
1128 GB ≥ 892.8 GB required, making this the smallest true configuration in our house that fits the model. A full 8-GPU NVSwitch domain — all 8 GPUs' memory is addressable as one 1,128 GB pool.
DeepSeek V4 Pro
1.6 trillion parameters. Even a full 8-GPU H200 node (1,128 GB) can't hold it — this is a rack-scale-only model.
What it means: parameter count is the model's size. “1600B” means 1600 billion weights. Even for a Mixture-of-Experts model where only 49B are active per token, every expert must still reside in memory at once — one cannot decline to load the experts not needed this particular token.
20480 GB ≥ 1920.0 GB required, making this the smallest true configuration in our house that fits the model. The largest real cluster that exists today: 72 Blackwell Ultra GPUs across the rack share one 20 TB NVLink memory pool.