An independent benchmark of local and hosted models on real coding tasks, run on my own machine and written up as I work things out. The posts make sense of the numbers. The numbers themselves all live in the data page.
Latest
19 July 2026 · same model, six harnesses
Wrap one fixed local model in six coding harnesses, hand them the same bug fixes, grade them the same way. The five that commit their edits score the same; what differs, by multiples, is time and tokens. The sixth writes fine code but often writes none.
Read the writeup →Where it started
29 June 2026 · local vs API
On this workload a local DeepSeek V4 finished real coding tasks faster in wall-clock than Sonnet or Opus over the API, at roughly Sonnet-grade quality, with the caveats left in rather than sanded off.
Read the writeup →24 June 2026· updated 4 August 2026 · the first post
Most local models feel quick at 8k and turn to dial-up by 150k, it comes down to how they handle attention, and chasing the few that stay fast is where all of this began. (Aug 2026: a context-cap bug invalidated the quality half — see the note up top.)
Read the writeup →The data
The full interactive scorecard behind these writeups, with the exact prompts, the per-task detail, and every run.