A KV cache is quantized to three bits with no accuracy loss

Claim

Google Research introduced TurboQuant, quantizing the KV cache to three bits without training or accuracy loss for up to eightfold performance, while Yann LeCun and colleagues unveiled the first world model training stably end-to-end from raw pixels and planning up to 48 times faster.

Reported metric: 3 bits / 48x

Reporting date: 2026-03-25 (not necessarily the event date)

Sources

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

• Welcome to March 25, 2026 (theinnermostloop.substack.com)