Sarvam 105B Explained: Benchmarks, Price and How to Use It

Sarvam 105B is India's flagship open-weight language model, a Mixture-of-Experts reasoning model from Bengaluru's Sarvam AI released under the Apache 2.0 licence. At its Epoch event, Sarvam shipped updated variants for voice agents and work agents. Here is what it is, how it scores, what it costs and how to use it.

Quick summary

• 105B+ total parameters, 10.3B active per token (Mixture-of-Experts), per Hugging Face
• Trained from scratch on 12 trillion tokens, 128K context window
• Open weights under Apache 2.0 on Hugging Face and AIKosh
• API price: ₹15 per 1M input and ₹60 per 1M output tokens
• Epoch update adds a conversations variant and a tuned work-agent variant

What is Sarvam 105B?

Sarvam says the model was trained entirely in India on compute from the IndiaAI mission, using in-house datasets across pre-training, fine-tuning and reinforcement learning. It uses Multi-head Latent Attention to keep long-context inference efficient. It powers Indus, Sarvam's AI assistant. The first checkpoint arrived in February and the weights were opened in March, according to Sarvam's blog and The Indian Express.

How good are the benchmarks?

Sarvam's documentation lists 98.6 on Math500, 88.3 on AIME 25 (96.7 with tools) and 49.5 on BrowseComp, and claims it wins 90% of pairwise comparisons on Indian-language tasks. These are the company's own numbers. Independent tracker benchgap lists measured scores of 78.7% on GPQA Diamond and 84.8% on IFEval, so it is strong for its size but not a frontier-class model on every test.

For voice-call workloads, Sarvam published a production benchmark comparing its conversations model with rival small models:

Treat this as vendor-run: it uses Sarvam's own data and was not independently reproduced.

What did Sarvam announce at Epoch?

Two updated variants: one tuned for low-latency conversations, one reasoning variant for agent harnesses. For work agents, Sarvam says routing is the useful trick: in its own setup 36% of tasks go to Sarvam 105B and 64% to the much larger GLM 5.2, cutting serving cost by roughly 40%. It also launched Voice Agents at ₹3.50 per minute and an early beta of Sarvam Code.

How much does Sarvam 105B cost in India?

Reasoning tokens are billed as output. For comparison, Sarvam lists Gemma 4 31B at ₹35 / ₹90 and GLM-5.3 at ₹126 / ₹390 on the same platform.

How do I use Sarvam 105B?

• Chat: use the Indus assistant from Sarvam.
• API: create a key in the Sarvam dashboard, point an OpenAI-compatible client at https://api.sarvam.ai/v2 and set the model to sarvam-105b. Reasoning defaults to low and can be turned off with reasoning_effort set to none.
• Self-host: download sarvamai/sarvam-105b from Hugging Face or AIKosh. A 105B model needs serious GPU memory, so most people will use the API.

Is it worth using?

If your work involves Hindi, Tamil, Marathi or code-mixed text, it is the most India-focused open model available, and the pricing is low. For pure coding or hard reasoning, bigger models still lead. See our look at Sarvam Prism and the open model comparison for context.

Related articles

Sarvam Prism Explained: India's 44 ms Open Decision ModelTencent Hy4 Launched: Vision, Computer Use, Price ExplainedAnthropic FlashProve Explained: Proving Which AI AnsweredMistral Large 4 Open Weights: Benchmarks and What It Needs