VoxCPM2
Speechopenbmb/voxcpm2OpenBMB's tokenizer free, diffusion autoregressive text to speech model, Apache-2.0 and free for commercial use. Its model card states 2B parameters, 30 languages and 48 kHz output. We run it on a single L40S and charge the card time a narration actually takes.
Playground
Write a prompt, set the request, and read what it will cost before anything runs.
Nothing runs from this page. The toolkit and the API are the same endpoint, the same price and the same balance.
Pricing
Charged per million characters of text, with a minimum per reading that covers loading the model on short passages.
- Rate
- $26.50/M characters
- Minimum per render
- $0.003
- Charged
- When the job is accepted, not when it finishes
- Failed renders
- Refunded in full, automatically
- Cold starts
- Never billed to you
The price is taken when the job is accepted, at the rate published here. Waking a machine takes time, never money: cold starts are on us.
Performance
Two numbers matter: how long a render takes on a machine that is already up, and how long the first one takes when none is.
Machines sleep when nobody is using them, so the first call after a quiet spell waits for one to wake. That wait costs you time and never money: the price is the catalogue rate either way.
- Answer
- Immediate. The call returns an id and a
queuedstatus. - Delivery
- Poll
GET /v1/renders/{id}, or setwebhook_urland get a signed call when it lands. - Output link
- Signed, and valid for 24 hours
Hardware
The weights run on GPUs we rent, with no inference provider in between.
- GPU
- L40S
- Precision
- As published by the author
- Parameters
- 2B
Parameters
Every field this model accepts. Anything it would ignore is not listed.
modelrequired- stringUse "openbmb/voxcpm2".
promptrequired- stringThe text the voice reads out.
style- stringA direction, such as slower or warmer.
seed- integerSame seed, same render. Random without it.
webhook_url- stringWe sign a POST to it when the render lands.
API
One endpoint serves every family. The model field picks the machine, and the rest of the body follows the model.
curl -s -X POST https://klemet.app/api/v1/renders \ -H "Authorization: Bearer $KLEMET_KEY" \ -H "content-type: application/json" \ -d '{"model":"openbmb/voxcpm2","prompt":"Your render is ready."}'It answers 202 at once, before the render starts:
{ "id": "rnd_3f8c21d0", "object": "render", "kind": "speech", "model": "openbmb/voxcpm2", "status": "queued", "created_at": "2026-09-14T09:12:44.000Z", "finished_at": null, "output": null, "estimated_usd": 0.003, "usd": null, "error": null}Read it back with GET /v1/renders/{id} until the status is succeeded or failed. A successful render carries an output.url. Create a key in your account, and the whole reference is in the docs.
Limits
What one call may ask for, and how many calls you may make.
- Per render
- Up to 40,000 characters per request, 48 kHz output
- Renders
- 20 per minute, 300 per hour
- Reads
- 120 per minute, 2,000 per hour
- Over the limit
429, with the window and the reset in the headers
Weights and licence
These are open weights. You can read the licence and run them yourself.
- Published by
- OpenBMB
- Weights
- openbmb/voxcpm2
- Licence
- apache-2.0
- Access
- Open: the weights download without an agreement.
The licence is the author's, not ours. What you may do with what you generate is written there, and it is worth reading before you ship.
FAQ
What am I charged if a render fails?
Nothing. The price is taken when the job is accepted and returned in full if the engine cannot finish it. The refund shows up in your activity as its own line.
Do you train on my prompts or my outputs?
No. Prompts and outputs are stored so you can read a render back and download it; they are not used to train anything, and the output links we hand out are signed and expire.
Why is the first render slower?
Machines sleep when nobody is using them. Waking one for VoxCPM2 and loading its weights took 2 min 16 s at our last measurement, against 3.3 s once it is up. The wait is yours; the extra card time is ours.
Can I use what I generate commercially?
That is the author's call, not ours. VoxCPM2 is published under apache-2.0, and the terms are in the licence itself.
Can I run these weights myself?
Yes. The weights download without an agreement, and nothing here is a wrapper around a closed API. We charge for the machines and the plumbing, not for access.
Explore
Other speech models running here.
This is the only one in its family for now.