Nvidia study maps KV caches between compatible LLMs

Nvidia study maps KV caches between compatible LLMs

Image: Primary

Published Friday, August 21, 2026 · 1:07 PM EDT

Researchers at Nvidia reported a method for transferring an LLM's KV cache between compatible models, avoiding a fresh prefill when an agentic workflow switches model sizes. In tests across matched-KV model pairs, the linear mapper ran 2.7 to 25 times faster than re-prefilling and retained 73% to 98% of target-model accuracy for four of six pairs.

Sources

In this story

Published by Tech & Business, a media brand covering technology and business. This story was sourced from VentureBeat and reviewed by the T&B editorial agent team.

Back to Newswire

Keep reading

AI Science

AI Science

Dust researchers report competitive transformer training without backpropagation

Researchers at qlabs.sh report that Dust, a method for training transformer language models without backpropagation, matched and sometimes exceeded that standard training method in experiments using large populations of trials. Ac...

Science Robotics

Science Robotics

SimForcing researchers report gains in robot video prediction and policy learning

Researchers report in a preprint that SimForcing, a system for predicting robot videos from robot actions, outperformed the compared methods on four video-quality measures on the Bridge dataset without external embodied pretrainin...

AI Capital

AI Capital

Ghost announces $11 million seed round and opens Core computer preorders

Ghost announced an $11 million seed round led by Andreessen Horowitz and opened preorders Monday for Core, its $3,499 computer designed to run AI agents that take actions on a user's behalf. Abstract, Audacious Ventures, SV Angel ...

AI Products

AI Products

OpenAI seeks UAE investors to anchor $30 billion financing

OpenAI is in talks with multiple investment funds from the United Arab Emirates to help anchor a $30 billion financing round, Bloomberg reported, citing people familiar with the matter. The funds include Abu Dhabi-based MGX; the f...

AI Infrastructure

AI Infrastructure

Reflection introduces Beam and plans open-weight release this month

Reflection introduced Beam, a model for coding, reasoning and tasks that use tools, and said it will release the model's weights and developer artifacts later this month. The model is undergoing final safety testing and evaluation...

Security AI

Security AI

AWS reports 89.0% success for Continuum on code security benchmark

AWS says its Continuum security system passed 819 of 920 tasks on CyberGym-E2E within the benchmark's 90-minute limit, achieving an 89.0% success rate. That exceeds the previous public high of 65.9% by 23.1 percentage points. The...