DeepSeek V3.2, not served here yet
DeepSeek V3.2 pairs DeepSeek Sparse Attention for long contexts with reasoning that runs inside tool calls, under an MIT license. OpenRelay does not serve it yet; tell us if you want it.
- Text in, text out
- Not served yet
- No public rate
Not served yet.
GET /v1/models lists DeepSeek V3.2 with status request_access, but nothing serves it: there is no backend behind the id and no public rate. A request tells us someone wants it.
openrelay/deepseek-v3.2Tell us you want DeepSeek V3.2
Leave a work email. We email you if we start serving it, and nothing else goes to this list.
The spec sheet.
What DeepSeek has published about the model, with the source on each row.
- On OpenRelay
- openrelay/deepseek-v3.2: Text in, text out. Listed with status request_access, not served. GET /v1/models
- Parameters
- 685B in the published checkpoint DeepSeek model card
- Attention
- DeepSeek Sparse Attention (DSA), built to cut compute on long contexts DeepSeek model card
- Agentic training
- A large-scale agentic task synthesis pipeline, to bring reasoning into tool use DeepSeek model card
- License
- MIT, open weights DeepSeek model card
DeepSeek V3.2, answered.
Can I call DeepSeek V3.2 through OpenRelay?
Not today. openrelay/deepseek-v3.2 appears on GET /v1/models with status request_access, but nothing serves it: a chat completions call returns 403 model_access_required. The form on this page tells us you want it.
Is there a DeepSeek model with a public rate?
Yes. DeepSeek V4-Flash is priced per token as openrelay/deepseek-v4-flash, and its page has the rate and a quickstart.