mojojojo
One mojojojo credit is $0.001 and buys milliseconds here.
| What | Rate | Works out as | 1000 credits ($1) buys |
|---|---|---|---|
| Browser mojojojo.js, WebAssembly |
free, unlimited | runs on the caller's machine | everything |
| CPU one core, 2 GB |
1 credit / 10000 ms | $0.36 / core-hour | 2.8 core-hours |
| GPU RTX 5090, time-sliced |
1 credit / 2000 ms | $1.80 / GPU-hour | 33 GPU-minutes |
Each project uses the bundled folded static 512-dimensional model. Source documents stay durable in mojojojo Postgres; portable int8 indexes are rebuilt or moved between workers without changing your project ID.
per 1,000 queries
Mojo int8 exact retrieval. No idle-instance fee. The response includes measured worker latency.
per 1,000 queries
cuVS CAGRA routing only when the worker passes a real build, self-retrieval, and latency probe; otherwise it stays exact.
250,000 documents per project
Batch upserts, metadata, durable source documents, automatic replica hydration, and API-key isolation included.
Dedicated retrieval capacity starts at $799/month with an agreed vector and latency envelope. Shared usage remains metered against the same credits as compute, with fractional charges carried forward rather than rounded per request.
Per account, for deployed sites, models, and user-uploaded assets.
Only storage above the included gigabyte. Usage is visible in Cloud.
Public assets are served from mojojojo's global static domain with no delivery surcharge.
The clock starts when your program starts. Interpreter startup, the sandbox, and the process fork are ours. Warm slots make it about 60ms end to end anyway.
Your imports are read out of your code and installed by uv into a cached environment. First time for a package set: a second or two. Billed: nothing.
Transpiling your kernels to Mojo and compiling them happens on our machines, on our clock, after your run has returned.
Every new mojojojo account gets 1000 credits — $1 of compute the first time it runs something here. No card, no trial clock. At the CPU rate that is nearly three core-hours, which is a lot of backtests.
One grant per account, ever. Server-side runs need an account for exactly this reason: free compute behind an anonymous endpoint is free compute per request, and it would be strip-mined within a day. The browser runtime needs no account because it costs us nothing.
| Job | Time | Cost |
|---|---|---|
| A 40ms data transform, called 1000 times | 40 s | 4 credits — $0.004 |
| A backtest that takes 3 s (native, after the first run) | 3 s | 0.3 credits — $0.0003 |
| An hour of continuous CPU work | 60 min | 360 credits — $0.36 |
| Ten minutes of GPU | 10 min | 300 credits — $0.30 |
Sub-credit runs are not rounded up. Milliseconds accrue against your balance and a credit moves when they cross the rate, so a thousand 40ms calls cost the same as one 40-second call.
Demand above what this box can serve spills over to rented capacity — RunPod for GPUs, Hetzner for CPU — and only when paid demand covers the bill. The policy is public: see the docs. Prices above do not change when that happens; a queue does, and queue time is never billed.