An MIT-licensed model folds 1.6 trillion parameters into a tenth the KV cache

Claim

DeepSeek's MIT-licensed V4-Pro-DSpark folds 1.6 trillion parameters and a million-token context into speculative decoding that sips a tenth the KV cache, as a field guide crowned DeepSeek, GLM 5.2, MiniMax M3 and Nemotron 3 Ultra frontier-class coders holding the same three-to-six-month gap they have kept for eighteen months.

Reported metric: 1.6T params / 10% KV cache

Reporting date: 2026-06-28 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

• Welcome to June 28, 2026 (theinnermostloop.substack.com)