Archive

Benchmarks

8 stories filed under Benchmarks.

1. Wednesday, 4 November 2026

   An M1 Max on Omarchy beat macOS on GPU memory bandwidth

   Joshua Warren measured 368 GB/s of GPU memory bandwidth on an M1 Max running Omarchy, against 361 GB/s on macOS and 241 GB/s before, in his post. He said one bit fixed it.

2. Wednesday, 4 November 2026 · In brief

   A Thunderbolt RTX 5060 Ti decodes Qwen3.8-27B at 24 tokens/s on an M1 Max MacBook Pro running Linux

   Joshua Warren plugged an RTX 5060 Ti, in a Thunderbolt box, into an M1 Max MacBook Pro running Linux, and in his post “Qwen3.8-27B” decodes at 24 tokens/s, 2.6x faster than the laptop's own GPU.

3. Tuesday, 3 November 2026 · In brief

   Nemotron 3.5 Bolt ran at 90 tps on an Nvidia N1X under Omarchy

   On Omarchy, on an Nvidia N1X, David Heinemeier Hansson ran Nemotron 3.5 Bolt at 90 tps, diagnosing a sleep issue while offline.

4. Tuesday, 27 October 2026 · In brief

   An AMD RX 9070 on an M2 Max reports a Qwen rate with one Mesa setting

   Joshua Warren said Qwen generates at about 33 tok/s with RADV_PERFTEST=dmashaders, up from 25, on an AMD RX 9070 (16 GB) over Thunderbolt on an M2 Max running Omarchy, on the same greedy output, in this post.

5. Saturday, 17 October 2026

   Muse Spark Open is reported on old Omarchy laptops

   Joshua Warren said he ran all 816B parameters of Muse Spark Open, at 2-bit, on three old M1 and M2 laptops running Omarchy and a 2022 Mac Studio on macOS, in this post. He linked an omarchy-cluster receipt.

6. Thursday, 15 October 2026 · In brief

   An M2 Max test reports a faster mixture-of-experts decode

   Joshua Warren said mixture-of-experts models got 20% faster on Apple Silicon running Linux, with gemma-4-26B moving from 20.4 to 24.6 tok/s on an M2 Max, in pull request 180.

7. Thursday, 8 October 2026 · In brief

   A decision-model test on Omarchy QA screenshots is called faster and more accurate

   ThePrimeagen said he ran OpenAI's new decision model on the previous day's Omarchy QA results, with the same prompt and image. He said it was generally much faster and that it was more accurate at finding things to click.

8. Saturday, 3 October 2026 · Corrected

   An M1 benchmark under Omarchy puts oMLX ahead of Apple's MLX

   Joshua Warren, of the Omarchy M team, said oMLX reaches 39.7 tokens a second on an M1 MacBook Pro running Omarchy, against 27.3 on Apple's MLX, in his benchmark.