Key takeaways

  • Sarvam 105B is a mixture-of-experts model with 106 billion parameters, 10.3 billion active per token and a 128K-token context window.
  • An updated checkpoint, Sarvam-105B-1711, went up on Hugging Face under Apache 2.0 on November 17, 2026, and the sarvam-105b API now runs on it.
  • The API costs ₹10 per million input tokens and ₹25 per million output tokens at introductory rates through the end of 2026, and new users get ₹100 in free credits.

The main Sarvam AI model today is Sarvam 105B, an open-weight reasoning model from the Bengaluru company with 105 billion parameters, of which about 10.3 billion run for each word. On November 17, 2026, Sarvam put an updated checkpoint, Sarvam-105B-1711, on Hugging Face under the Apache 2.0 license, and its API docs say the sarvam-105b API now runs on those weights. Through the end of 2026 the API costs ₹10 per million input tokens and ₹25 per million output tokens.

Sarvam trains its models from scratch, for Indian languages first. The new checkpoint is aimed at agents: Sarvam's own tests show its biggest gains in following instructions and calling tools.

Which models Sarvam offers

Sarvam released two language models in early 2026: Sarvam 30B and Sarvam 105B. Both are mixture-of-experts models, which split the network into 128 "experts" per layer and run only a few of them for each token. Both were trained in India.

Model Total / active parameters Context window Weights API status (as of November 22, 2026)
Sarvam 105B (now the 1711 checkpoint) 106B / 10.3B 128K tokens Apache 2.0, on Hugging Face Live as sarvam-105b
Sarvam 105B Conversations Same weights, post-trained for dialogue 32K tokens Not a separate download Live as sarvam-105b-conversations
Sarvam 30B 32B / 2.4B 64K tokens Apache 2.0, on Hugging Face Deprecated in the API
Sarvam Prism Built on Sarvam 105B 32,768 tokens per question On Hugging Face as sarvamai/prism Live as sarvam-prism
Sarvam-M 24B On Hugging Face Deprecated, no longer served

Parameter counts are from Hugging Face; context windows from Sarvam's docs.

  • Sarvam 105B was pre-trained on 12 trillion tokens. It reasons before answering by default; you can turn that off per request.
  • Sarvam 105B Conversations is the same model tuned for voice agents and chatbots. Reasoning is always off, and the context window is a quarter of the size.
  • Sarvam 30B was pre-trained on 16 trillion tokens and Sarvam tuned it for L40S GPUs and, compressed, a MacBook Pro M3. The docs tell API users to migrate to 105B, but the weights stay on Hugging Face.
  • Sarvam Prism doesn't write text. It returns a probability for each option of yes/no, pick-one and score questions, for routing, triage and moderation.

On languages, Sarvam's blog says training covered all 22 scheduled Indian languages. The API docs promise less: the 10 most-spoken Indian languages plus English, in native script, romanized or mixed with English.

What is new this month: Sarvam-105B-1711

Two repositories appeared on Sarvam's Hugging Face page with an "Initial release" commit on November 17, 2026: sarvam-105b-1711-bf16, the full-precision weights, and sarvam-105b-1711, a build compressed to NVIDIA's 4-bit NVFP4 format. That is why Hugging Face lists the second at 56B: the file is smaller, not the model.

The model card says 1711 uses the same base model as Sarvam 105B and was post-trained further for four things: tool calling, instruction following, agentic tasks and Indian languages. The card doesn't explain the name. It matches the release date written day-first, 17/11, but Sarvam hasn't said that.

The docs say both sarvam-105b and sarvam-105b-conversations now run on the 1711 weights, with no code changes needed.

Not confirmed by any Sarvam page: a launch post for 1711, the date the API switched, or any Sarvam comment on Chinese open models. Sarvam's Epoch recap mentions updated 105B variants for conversational and work agents but gives no date and doesn't name 1711.

Is Sarvam AI open source

The weights are open. Sarvam 105B, the 1711 checkpoint and Sarvam 30B are all published under Apache 2.0, a permissive license that allows commercial use, fine-tuning and redistribution. The 1711 card shows how to serve it with vLLM on two GPUs at the full 131,072-token context.

The training data isn't published, so "open weights" is the accurate term rather than open source in the strict sense.

How much the Sarvam API costs

Prices are from Sarvam's pricing page as of November 22, 2026. Rupees per million tokens, excluding GST, same on every plan. Reasoning tokens bill as output.

Model Input Cached input Output
Sarvam 105B (sarvam-105b) ₹10 ₹0.25 ₹25
Sarvam 105B Chat (sarvam-105b-conversations) ₹10 ₹0.25 ₹25
Sarvam Prism (sarvam-prism) ₹4 ₹0

The 105B rates are marked "introductory pricing" through the end of 2026,. At Epoch, Sarvam put it at $0.80 per million blended tokens, against its own figures of $4.50 for GPT 5.4 Mini and $9.00 for Gemini 3.5 Flash.

Is Sarvam AI free

Partly. The weights are free to run on your own hardware. The API gives verified new users ₹100 in credits that never expire, then charges per token. Pay-as-you-go has no monthly fee; Pro costs ₹9,999 a month and Growth ₹49,999, mainly for higher limits (60, 200 and 300 requests a minute on Sarvam 105B).

Reasoning tokens count toward max_tokens, which defaults to 2,048, so set it higher or the model may think through the whole budget. Other providers' rates are in our AI API pricing table.

How Sarvam 105B compares

Sarvam's own tests of 1711. The model card compares it with models of similar cost, all run by Sarvam. Selected rows:

Benchmark Sarvam-105B-1711 Gemma-4-31B Nemotron-Super-120B GPT-5.6 Luna Claude Haiku 4.5 Gemini 3.5 Flash-Lite
IFEval (instruction following) 96.5 94.7 90.8 93.1 86.1 95.6
τ²-Bench (tool use) 85.5 81.1 77.5 79.5 62.1 83.1
Indic Suite (Sarvam's internal) 79.1 78.6 39.5 82.4 61.3 79.7
GPQA Diamond (science) 80.4 84.3 78.9 88.7 67.9 85.2
AIME 2026 (math) 89.7 89.0 87.6 97.2 86.3 83.2
LiveCodeBench v6 (coding) 75.3 81.9 75.5 85.0 57.0 73.9
BrowseComp (web research) 41.4 33.2 27.9 76.3 18.4 42.9

Scores are for the full-precision weights. Sarvam leads on following instructions and calling tools, and sits within about three points of the top Indian-language score. GPT-5.6 Luna leads on knowledge, math, coding and search, and the much smaller Gemma 4 31B beats it on science and coding.

Against the original February checkpoint, as reported in Sarvam's blog, IFEval rose from 84.8 to 96.5 and τ²-Bench from 68.3 to 85.5, while BrowseComp fell from 49.5 to 41.4. The settings may differ, so read that as a direction.

Independent tests. Artificial Analysis scored the original Sarvam 105B at 18 on its Intelligence Index in April 2026, and Sarvam 30B at 12. In the same write-up gpt-oss-120B (high) scored 33 and GLM-4.5-Air 23. There is no score for 1711 yet.

That independent number matters more than any vendor table: the February model sat well behind open rivals of its size. The 1711 update targets the agent skills where that gap hurt, but no outside lab has checked it yet.

Who should use it

Sarvam 105B fits products that talk to Indian users in their own languages, like voice agents and support bots, and teams that need weights they can run on their own servers. For English coding, math or research agents, Sarvam's own table shows other models doing better. Watch for an Artificial Analysis score for 1711, and for the API price after the introductory rate ends on December 31, 2026.

Frequently asked questions

Is Sarvam AI open source?

Sarvam 105B, the new 105B-1711 checkpoint and Sarvam 30B are open-weight models under the Apache 2.0 license, which allows commercial use. Sarvam has not released its training data.

Is Sarvam AI free?

The weights are free to download and run yourself. The API gives verified new users ₹100 in free credits, then charges per token, with no monthly fee on the pay-as-you-go plan.

What is Sarvam's new model?

Sarvam-105B-1711, released on Hugging Face on November 17, 2026. It is Sarvam 105B post-trained further for tool calling, instruction following, agentic tasks and Indian languages.

How does Sarvam AI compare with ChatGPT?

In Sarvam's own tests, Sarvam-105B-1711 beats OpenAI's GPT-5.6 Luna on instruction following and tool use, but GPT-5.6 Luna leads on knowledge, math, coding and web research, and slightly on Sarvam's Indian-language suite.

Sources

  1. Sarvam-105B-1711 (BF16) model cardSarvam AI on Hugging Face
  2. Sarvam-105B-1711 (NVFP4) model cardSarvam AI on Hugging Face
  3. Sarvam-105B model docsSarvam AI
  4. Pricing, credits and rate limitsSarvam AI
  5. Open-sourcing Sarvam 30B and 105BSarvam AI
  6. Sarvam-30B (deprecated) docsSarvam AI
  7. Sarvam Prism docsSarvam AI
  8. Sarvam Epoch summarySarvam AI
  9. Sarvam models on Hugging FaceHugging Face
  10. Sarvam 105B and Sarvam 30B: everything you need to knowArtificial Analysis