Sarvam 105B
Sarvam 105B is a mixture-of-experts model with 10.3B active parameters. It uses Multi-head Latent Attention and is published under Apache 2.0. Hosted access includes OpenAI-compatible chat completions.
Budget for reasoning output
Reasoning is enabled by default at low effort. Those tokens count toward the completion allowance, so a small output budget can leave no visible answer. The API guide documents this behavior explicitly.Sarvam 105B and Conversations API guide
Practical use
Practical guidance. Compare native-script and romanized versions of your real prompts. Track whether tool arguments and final answers remain correct in both forms. Published multilingual coverage does not guarantee equal quality for every domain.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | sarvam-105b |
| Expert routing | Eight routed experts from 128, plus one shared expertSarvam 105B official weights |
| Published context | 128K; developer notationSarvam 105B and Conversations API guide |
Related models
- Sarvam 105B Conversations
Sarvam AI · Sarvam
A post-trained Sarvam 105B variant for colloquial Indic dialogue, voice agents and multi-turn chat.
Launch date not confirmedAPI - Sarvam 30B
Sarvam AI · Sarvam
A 30B mixture-of-experts language model with downloadable Apache 2.0 weights.
Launch date not confirmedDeprecated API; check access detailsDownloadable - Sarvam Prism
Sarvam AI · Prism
A Sarvam 105B-based decision model that assigns probabilities to predefined choices.
Launch date not confirmedAPI · Downloadable