Sarvam 30B
Sarvam 30B remains available as Apache 2.0 weights. Its API page is now a deprecation notice, with older endpoint details retained only for reference. This entry confirms weight access, not continued hosted service.
Deprecated API; check access details
The smaller model uses a different attention design
The card describes grouped key-value attention and fewer layers for efficient inference. It is not simply the 105B MLA model with a lower parameter count.Sarvam 30B official weights
Practical use
Practical guidance. Compare local deployment against the current hosted flagship using actual message lengths and concurrency. The active count does not describe the full memory needed to store expert weights.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Deprecated API; check access details |
| Access | Downloadable |
| Model identifiers | sarvam-30b |
| Active parameters | 2.4B non-embedding active parametersSarvam 30B official weights |
| Architecture | 19 layers, 128 routed experts, top-six routing and a shared expertSarvam 30B official weights |
Related models
- Sarvam 105B
Sarvam AI · Sarvam
A 105B mixture-of-experts reasoning model for Indian-language text, code and tool calls, with 10.3B active parameters.
Launch date not confirmedAPI · Downloadable Sarvam M is a text-only multilingual model built on Mistral Small.
Launch date not confirmedDeprecated API; check access detailsDownloadable