Skip to content
Writing Benchmarks About
See the bench → Read the writing

Latest writing

01

How this site is built

Astro, static HTML, build-time SVG charts, no trackers, and ~6 KB of vanilla JS. A colophon, with reasons.

Read more → Jul 2026

Browse the full archive →

Browse by topic

  • Meta1

Start here

  • →How this site is built

The bench

Local LLM throughput

Tokens/sec on my own hardware, with methodology and run logs included.

  • Qwen3-32B AWQ · decode 75 tok/s
  • + n-gram spec · code edit 350 tok/s
  • 16-way concurrent, aggregate 656 tok/s

Qwen3-32B-AWQ · vLLM · RTX 5090 · full chart on the bench page

Want new posts as they land?

Grab the RSS feed: every post, the moment it ships, no inbox required. Email subscriptions are coming once the newsletter hookup goes live.

Follow via RSS

Hope you're having a lovely day!

New writing lands in the RSS feed the moment it ships. Email subscriptions arrive with the newsletter hookup.

rss.xml has everything

Engineer. Product work by day, a benchmarking habit by night. Everything here is measured on hardware I own.

Sections

  • Writing
  • Benchmarks
  • About

Elsewhere

  • Email
  • RSS

© 2026 Joe Guglielmucci. Hand-built with Astro: static HTML, charts are build-time SVG, no trackers, no chart libraries. The interactive layer is ~6 KB of vanilla JS. View source; it’s readable.

Psst, enjoying this?

New posts land in the feed the moment they ship, no algorithm in between.

Follow via RSS
  • →Home
  • →Writing
  • →Benchmarks
  • →About
  • →RSS feed
  • ⧉Copy page URL