Skip to content
Atlas Blog
LatestEngineeringBenchmarksDesign
atlasinference.io

Benchmarks

What we measured, on what hardware, with the harness attached.

All EngineeringBenchmarksReleasesDesign
  • Sep 1, 2026

    DFLASH-2: the fastest single-machine numbers Atlas has produced

    66.6 tokens per second on a stock build, one DGX Spark, one stream, and every figure reproducible from a commit.

    4 min · RS
Atlas Inference Engine

Zero-trust inference on hardware you own. Pure Rust and CUDA, built in North Carolina.

Blog

  • Latest
  • Engineering
  • Benchmarks
  • RSS feed

Atlas

  • atlasinference.io
  • Documentation
  • Benchmarks
  • Download

Community

  • GitHub
  • Discord
  • X
© 2026 Atlas Inference · Community Edition AGPLv3
blog.atlasinference.io