How to read this
Identical prompts, identical history, no system-prompt coaching. Times are measured server-side per model, from request to completed answer; token counts come from the provider’s own usage report.
Ask one question and watch four open-weight models answer it at the same time — same prompt, same conditions, no cherry-picking. Every column shows the model’s own words, how long it took, and how many tokens it used. Follow-ups continue each model’s own thread, so you can push on an answer and watch where the reasoning holds — and where it folds.
One prompt goes to all four models at once. After the first round, ask a follow-up and each model answers in its own thread, with its own history. Nothing is stored — the conversation lives only in this page.
One quick human check per question — it keeps bots from draining the Arena’s answer budget.
Identical prompts, identical history, no system-prompt coaching. Times are measured server-side per model, from request to completed answer; token counts come from the provider’s own usage report.
Elapsed time mixes model speed with provider load and routing on the Hugging Face router, so treat it as a snapshot of this run — not a benchmark. The leaderboard discipline applies here too: one run is an anecdote, not a ranking.
GPT, Claude, and Gemini answer side by side in the Model Arena, on the same rules as here. This Open Arena stays as the open-weight lane — the models here can be downloaded and run by anyone.