JevBench · ImageJevBench · AudioJevBench
Submit a model for evaluation
Hand in a model and we will run it through our benchmarks and publish the result. Anyone can submit; no account needed.
How it works
- The regular queue is free. Submissions are evaluated in the order we receive them.
- Top-10 models are re-evaluated with every release; the rest of the leaderboard is re-evaluated less often.
- Fast lane: results within 48 hours of payment, or a full refund on request. It covers JevBench and ImageJevBench.
Fast-lane fees
JevBench is independent, volunteer-run and open source. GPU and token costs are real, and a fast-lane fee helps cover them.
Per benchmark for API-served models or open models up to about 9B parameters.
Per benchmark for larger open models that we run on our GPUs.
Choosing both JevBench and ImageJevBench charges the selected tier twice. Applicable tax is calculated at checkout.
What the fee does and does not change
- Payment buys earlier scheduling only. It never buys a different method, score or rank. Every model uses the same frozen task set and rules.
- We decide which models we evaluate and when. We may also evaluate any model early on our own if it is of public interest.
- Every submission goes through code review. If something needs to change, we tell you exactly what; the clock pauses and the full 48 hours start again when you send the update. We refuse and fully refund requests that are unsafe or cannot be evaluated fairly.
- If we miss the 48-hour deadline, you can ask for a full refund. The clock pauses while we wait for information requested from you.
- A public paid run carries a visible “priority run” marker in the data. A private report is never published without your team's consent.
Think you should be exempt? If you work for a non-profit or in research, tell us briefly why and we can waive the fee at our discretion. Email us about a fee waiver.
What we need
A name, how to reach the model (a GitHub link, a Hugging Face link or a public API URL, at least one), your email and the benchmarks you want. For an API model you can add a key; it is encrypted the moment we receive it and used only for this evaluation. Never paste keys anywhere else.
Hosted-endpoint timing (on request)
Self-hosted models are timed on our own hardware over loopback (on-GPU, no network), which is not directly comparable with API latency, so the ranking uses the same estimate for all of them: 2× measured + 150 ms for network. If you prefer, give us a hosted endpoint and we time it exactly like an API, network included, with no assumption. Mention it in the notes field below or email info@productivity-boost.com.
Your submission
We store what you enter here to run and report on the evaluation, and delete a submitted API key afterwards. See the privacy policy and terms.