TOKRACE is an LLM speed testing bench for developers and AI reviewers. It runs one prompt across multiple LLMs and compares TTFT, output TPS, peak speed and stability.

Is TOKRACE free to use?

You can run preset sample models for free. Advanced users can add their own model API keys. The project is open source and the server does not persist keys.

Will my API key be exposed?

API keys are stored only in your browser localStorage. Requests pass through the same-origin service only to avoid browser CORS; the server does not log or persist keys or conversation content.

Where does leaderboard data come from?

Speed and audience leaderboards use metrics and votes that users voluntarily share anonymously. They contain speed numbers only, not prompts or API keys.

Free samples · Real-time speed tests · Shareable evidence

A real-time model speed benchfor developers and AI reviewers

Run one prompt across multiple LLMs and compare TTFT · output TPS · peak speed · stability. Start with a free sample, then switch to your real task and model setup.

Run Sample ▶Professional Mode Templates Invite Rewards GitHub

ByAI拯救打工人|Open Source

Write a 5,000-word essay on why speed itself is a capability● LIVE

Model A 🥇312 tok/s

Model B 🥈187 tok/s

Model C 🥉96 tok/s

0.68s

TTFT

102

Thinking TPS

312

Output TPS

387

Peak tok/s

↑ Demo animation · real data comes from your arena runs

Run a sample

Try a real concurrent speed test with preset models without configuring keys.

Use your prompt

Test your real task, long-form input, code task or visual prompt against configured models.

Share evidence

Generate snapshots, voting pages and long images so readers can inspect the result.

Why test speed with TOKRACE

🏁

Concurrent tests with the same prompt

Send the same prompt to multiple models at the same time to reduce ordering, network and time-of-day bias.

⏱

Real-time TTFT / TPS comparison

Track TTFT, thinking TPS, output TPS, peak speed and token counts side by side.

📈

Zoomable speed curves

Inspect 2-second sliding-window tok/s curves, zoom into peak moments and export PNG evidence.

🔌

Free samples + bring your own key

New users can run preset models immediately; advanced users can connect OpenAI-compatible or native Anthropic endpoints.

🔐

Keys stay local

API keys stay in encrypted localStorage. Requests pass through the same-origin server once and are not stored.

📷

Built for publishing

Edit titles and notes, use screenshot mode, customize watermarks and copy metrics as Markdown.

🧪

Validate as you connect

Fetch provider model lists, test connectivity and configure private parameters per model.

🖼

Automatic SVG rendering

SVG code in model output is rendered as an image in a sandbox with scripts disabled.

Your keys stay yours

API keys are stored only in your browser localStorage. Requests pass through a same-origin proxy only to avoid browser CORS; the server does not log or persist keys or conversation content. Prefer no hosted proxy at all? The project is open source:git clone && npm run dev.

Run Sample ▶View source