RTX 3090 Hits High Long-Context Throughput with 33B Model

GreenGames · hn · 2026-07-11

A Hacker News benchmark share reveals that running Laguna XS 2.1 33B on an RTX 3090 achieves a peak speed of 296 tok/s, while maintaining 152 tok/s at a 256K context length.

The focus here isn't on the model's release, but rather on a concrete local deployment / long-context inference performance metric, making it highly relevant for those interested in single-GPU inference, throughput, and context costs.

Original post →

More from Infra

Infra channel →