Sarvam 105B
Sarvam 105B
Built for Indian languages first, and it is not a translation layer over an English model.
Updated
The read
Sarvam trained this for India: twenty-two languages treated as first-class rather than as an afterthought, with a tokeniser and a data mix to match, published as open weights. For Indic-language products it beats models several times its size, because the big labs simply do not allocate their training budget this way. In English it is mid-pack, and it is not trying to be anything else.
Where it shines
- Twenty-two Indian languages handled as first-class, not translated
- Tokeniser tuned for Indic scripts, which cuts cost per request
- Open weights, so it can be deployed inside India
Worth knowing first
- Mid-pack in English against models of similar size
- Short context window
Index score
74.3
Provisional. This score is read from the product's public capability, not from the one-prompt rebuild the fully reviewed entries went through, so treat it as a placing rather than a verdict.
- Rank
- 42 of 53
- Builds
- LLMs
- Category
- Open weights
- Output
- Published weights, Indic-language first
- Pricing
- Free tier
- Best for
- Products serving Indian languages properly
Scores are our own editorial judgement, weighted craft 55, speed 10, control 15 and value 20.
Also worth a look
Full index- 05Kimi K3Kimi K3ProvisionalThe largest open-weights model anyone has shipped, and it reads a million tokens without blinking.Open weightsFree tier87.5
- 06GLM-5.2GLM-5.2ProvisionalFrontier-class output, an MIT licence, and a million tokens of context. That combination is the story.Open weightsFree tier87.1
- 12Qwen 3.6Qwen 3.6ProvisionalThe open Qwen release most people should actually run, in sizes that fit real hardware.Open weightsFree tier83.3
