A reported training target is not a finished model specification. Ten trillion parameters is a striking number, but the conclusions it supports are much narrower than a capability ranking.
Keep the claim attributed
The Financial Times report of August 7, also listed on Yahoo, describes early training of a ByteDance model with as many as 10 trillion parameters, citing three people familiar with the work. Its accessible report excerpt says the exact size would be determined later. We have not independently verified the training run.
The full FT article requires a subscription. This update does not turn estimates about competitor model sizes into official disclosures, or infer a named accelerator fleet from a total parameter count.
A storage calculation with explicit assumptions
Our cover asks a deliberately limited question: if 10 trillion parameters were each stored in exactly two bytes, how much space would the weights occupy? The multiplication gives 20 trillion bytes, or 20 TB in decimal units. At an ideal packed four bits per parameter, it gives 5 TB.
Those figures exclude quantization metadata, caches, activations and runtime buffers. Training adds other state. They are hypothetical weight-storage calculations, not ByteDance’s memory requirements or a claim that either precision is used. A parameter count without a representation does not even give one unique byte total.
Stored parameters and active work differ
In a routed-expert design, only selected experts may participate for a token. A dense design uses a different pattern. Neither the architecture nor serving cost can be derived from the headline number alone. Communication, batching, sequence length and implementation also affect throughput.
Likewise, a larger parameter store is not a test result. To decide whether a released model is useful, compare tasks, correctness, latency and cost under documented conditions. A model can contain more weights and still fail a workflow that another system handles reliably.
The useful next evidence would be a named release or technical report with reproducible evaluations. Until then, preserve the distinction between reported ambition and demonstrated behavior. Repeating the largest estimate as a product fact only makes the original uncertainty harder for readers to see.
September 8: retain the 10T figure as attributed reporting, remove unverified chip counts and competitor sizes, separate hypothetical weight storage from active compute and capability.