Sarvam 1
Sarvam 1 is a text-completion base model. Its card explicitly requires downstream fine-tuning before chat or instruction use. It covers ten Indian languages and English under Sarvam's noncommercial license.
Research release
Do not mistake a base model for a chat model
The architecture has 28 layers and grouped-query attention. The page describes a 2B model, while the repository's automatic parameter display rounds to 3B; these labels should not be silently treated as identical measurements.Sarvam 1 official model card
Practical use
Practical guidance. Choose this checkpoint for adaptation or pretraining research. Compare a fine-tuned result against an instruction model on the same independent tasks. Raw completions are not a fair substitute for an assistant interface.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | sarvam-1 |
| Context window | 8,192 tokens Maximum position embeddings in the model card Sarvam 1 official model card |
| Training corpus | About 4T tokens, including 2T Indic tokensSarvam 1 official model card |
Related models
- OpenHathi 7B Hindi Base
Sarvam AI · OpenHathi
OpenHathi 7B Hindi Base is a Llama 2-derived model trained for Hindi, English and Hinglish.
Launch date not confirmedResearch releaseDownloadable - 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