The headline number on the newest open models looks impossible to run: hundreds of billions of parameters. The trick that makes them practical is the same one reshaping the whole field — sparsity. A flagship open model can carry roughly 744 billion parameters yet activate only around 40 billion for any given token, and pair that with a usable million-token context window.
Capacity you don't pay for on every token
In a dense model, every parameter fires for every token. A sparse Mixture-of-Experts model instead routes each token to a small subset of specialized experts. The full parameter count is where knowledge lives; the active count is what you pay to run. That decoupling is why a "744B" model can serve at roughly the cost of a much smaller dense one.
Total parameters are the library. Active parameters are the librarian you actually hire per question.
Why this is a big deal for open models
Sparsity is the great equalizer. It lets open releases match closed frontier capacity without demanding closed-frontier serving budgets. Combined with permissive licenses and long context windows, it puts genuinely capable systems within reach of teams that could never train one from scratch.
The catch
Sparse giants aren't free lunches. Serving them well is a systems problem: experts live across devices, so routing means heavy all-to-all communication that can dominate latency. Memory for the model and its long-context cache is substantial. And routing quality — sending each token to the expert that actually helps it — is where the real capability gap now lives.
For most teams the practical move isn't to run the biggest sparse model, but to understand that "parameter count" and "cost to run" have come apart — and to choose based on the second.
