Inference for openSystem One models
Typed answers with calibrated confidence, in milliseconds. Codiv serves OpenJev through a Jev-compatible API. Every account gets 100M input tokens free.
See it in action
Decisions, not tokens
A System One model reads your data once and answers every question with a probability, like a fast intuition you can measure.
Typed answers
Ask yes/no, choice and score questions. Every answer fits your schema by construction, because nothing is generated or parsed.
Calibrated confidence
Each answer is a full probability distribution with a confidence score, so thresholds actually mean something.
Milliseconds
One denoising step answers a whole set of questions. A typical request takes tens of milliseconds of model time.
Jev-compatible
The same wire API as TypeSafe's Jev. Their Python and TypeScript SDKs work unchanged; change the base URL.
Open all the way down
Open weights and an Apache-2.0 server. Anything you build on Codiv, you can also run on your own GPU.
Free to start
100M input tokens for every account while the experiment runs. No card required.
Playground
Try it in the browser
Typed questions, photos, chess, gomoku and chat, with every probability on screen. Sign in and run them on your free quota.
Any state and questions, with every request option.
Try itstate
Hi, I think I was charged twice for September. Could you take a look when you get a chance?
questions
- urgentnoulNeeds a reply within the hour?
- teamchoiceoutage · billing · feature
- tonescorecalm · annoyed · furious
one denoising step · 38 ms
A replay of typical answers. The playground runs the real API on your free quota.
Models
Open models, served fast
OpenJev 0.1
LiveDiffusionGemma 26B-A4B turned into a System One model: the answer template goes onto the diffusion canvas and one denoising step reads a distribution for every answer.
- Model id
openjev-latest- Question types
- noul · choice · score
- Base model
- DiffusionGemma 26B-A4B, NVFP4
- License
- Apache-2.0, weights and server
More System One models
Codiv is built to serve any open System One model. New models and sizes will be added as open diffusion models land.
Drop-in
Already using Jev? Change one line.
Codiv speaks the same API as TypeSafe's Jev, down to the error shapes and request ids. Point the official SDKs at Codiv and your code keeps working.
Compatibility notes →# Python or TypeScript: the official TypeSafe SDKs export TYPESAFE_BASE_URL=https://api.codiv.ai export TYPESAFE_API_KEY=sk-codiv-…
Free public experiment
Free for everyone
Free
Available now$0
No card required
- 100M System One tokens per account
- 10M text-generation tokens
- All models, including OpenJev
- 60 requests per minute per key
- Usage dashboard and up to 10 keys
Roadmap
OpenJev 0.1 public experiment
LiveDiffusionGemma 26B-A4B as a System One model, served free on api.codiv.ai.
Dashboard, keys & usage
LiveSign in with Google, GitHub or email. Track tokens per day.
Image input
LiveAsk typed questions about up to 8 images per request, alongside the state.
Text generation
LiveThe same DiffusionGemma at an OpenAI-compatible /v1/chat/completions, with its own free quota.
Playground
LiveTry reads, photos, chess, gomoku and chat in the browser, with every probability shown.
OpenJev 1.0
PlannedOnce vLLM PR #57250 is merged upstream. openjev-latest will move to it.
Batch mode
PlannedSend thousands of states in one call and collect the answers when they're ready. Built for backfills and nightly jobs.
More open System One models
PlannedNew models and sizes as open diffusion models land.
Make your first decision in five minutes
Sign up, create a key and call the API. It's free.