Can one box on a desk keep up with the cloud?

Eight open models ran on one ASUS GX10 that sits on a desk. Five paid cloud models ran the same nine jobs, with the same limits, and a judge that never saw a name. Here is who did the work best, who was fastest, and what it cost.

On the desk (one GX10) Cloud API

Best on the desk

–

Best in the cloud

–

Quality out of 100: the middle of 19 judged answers per model.

See all 13

The desk and the cloud share the middle of the table.

The top two are cloud models. Right behind them are two open models running at home, ahead of three paid cloud models. Pick any model to see where its points came from.

    So who wins each job? Not always the same model. See the nine jobs ↓

    No single model wins every job.

    Stronger color means a higher score. A ring marks the best score on each job.

    Score for each model on each job, out of 100

    Quality is half the story. The desk paid for its scores in hours. See the clock ↓

    The cloud finished in minutes. The desk took hours.

    Time to finish all 57 answers

    Log scale. Every model wrote the same 57 answers under the same length limits.

    Writing speed on the desk, tokens per second

      Tokens written per second once the answer starts. A token is about three quarters of a word. Green is DeepSeek split across two GX10s. It has not been judged yet, so it is not in the ranking.

      The desk is slow but it does not send a bill. The cloud does. See the bill ↓

      The best cloud score was not the most expensive one.

      Quality against the bill for one full run

      And the desk?

      No bill per answer. The GX10 draws wall power, and that has not been measured under load yet.

      An idle reading would flatter it, so there is no local dollar figure on this page. When the meter is on, it will appear here.

      What the desk gives you instead

      Your prompts never leave the room, and the model files (7 to 104 GiB here) are yours to keep.

      Numbers from a judge are one thing. Pictures are another. Look for yourself ↓

      What this run cannot tell you yet

      Long answers were cut off

      The ranked run capped answers at 512 to 4,096 tokens. Many long answers stopped mid-sentence and were scored as written.

      The judge reads, it does not run

      Grok 4.7 scored the text. It never opened a page or ran the code. At the old limit, no model's Mustang page showed a car on screen.

      Some cloud models think anyway

      Thinking was off where a model allowed it. GPT-6.1 Sol, Claude Sonnet 5.5 and Claude Opus 5.5 cannot turn it fully off, so they used their lowest setting.

      Not everyone is scored yet

      Gemini 4 Hydrogen and the two-GX10 DeepSeek have answers but no judge score. Power draw under load is not measured.

      The next run

      Room to finish (100,000 tokens for the picture jobs, 30,000 for the rest), checks that run the code and open the pages, a scored DeepSeek on two GX10s, Gemini judged, and a watt meter on the desk.

      How it was run