Unverified50% confidenceConceptExact time
When model parameter count is halved, the number of bytes that must be moved per forward pass is approximately halved, and given unchanged memory bandwidth, the number of tokens that can be generated per second nearly doubles.
1
Sources
50%
Confidence
Long-term
Relevance
11/22/2026
First Seen
Sources
Gemini 4 Neon Hands-On: Small Model Matches Flagship, Doubles the Speed
youtubeFAQtion11/20/2026
Related Entities
Related Claims
UnverifiedAfter offloading half of the model layers to system memory, the generation speed often drops sharply from over ten tokens per second to single digits.78% similarUnverifiedIn agentic tasks, the model repeatedly feeds context back in, causing input token consumption to balloon sharply, which is the key reason that estimating total cost by simply doubling the unit price becomes inaccurate.68% similarUnverifiedThe model's local throughput is approximately 9 to 22 tokens per second, consumes more tokens than the previous generation, and is slower when thinking mode is enabled.67% similarUnverifiedThe number of assignable hosts in a subnet equals 2 to the power of the number of host bits minus 2.65% similarUnverifiedThe measured VRAM of a resident service is 1.2–2.5× its weight size, with more stateful models exhibiting a higher multiplier.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1196269API
curl https://kongchang.com/api/v1/knowledge/claims/1196269MCP
get_claim(id=1196269)