Indian AI models: Sarvam 105B, Param2 and the GPUs to run them in India

Part of our AI models guide: 11 posts on choosing and running AI models.
India's own AI models have reached the point where a business can download one and use it for free. Sarvam AI's Sarvam 30B and Sarvam 105B are open weights under Apache 2.0, which allows commercial use; BharatGen's Param2 17B carries a non-commercial licence. All three are built for India's 22 scheduled languages, but on Artificial Analysis's independent index (April 2026, an earlier index version) Sarvam 105B still trails OpenAI's open gpt-oss-120B. By our estimate, Sarvam 105B runs on one 8-GPU H100, H200 or B200 server at full precision, and its compressed 8-bit (FP8) version fits on a single H200 or B200 GPU. Indian-language text can need several times as many tokens as English on older tokenizers, which raises API bills, and payment and government data must stay in India: a strong case for running these models on GPUs in India.
India's AI models at a glance
India now has a handful of open-weight models built for Indian languages: Sarvam AI, a startup, leads, and BharatGen, a programme led by IIT Bombay, is second. Open weights means you can download a model and run it on your own GPUs (see open-weight vs closed models). Size is counted in parameters, the numbers a model learned in training. Sarvam 30B, Sarvam 105B and Param2 are mixture-of-experts (MoE) models: only some parameters, the active ones, work on each token (the smallest unit of text an AI model reads and writes, often a short word or part of one), so they cost less to run than their size suggests. Context is how much text, in tokens, the model reads at once.
Model | Released | Size (total, active) | Languages | Licence | Notes |
|---|---|---|---|---|---|
Sarvam-1 | Oct 2024 | About 2.5B (named 2B) | 10 Indian languages and English | Sarvam non-commercial | Sarvam's first small model |
Sarvam-M | May 2025 | 24B | Indian languages and English | Apache 2.0 | Built on Mistral Small 3.1, a French open model |
Sarvam 30B | Feb 2026; weights 6 Mar 2026 | 32B total, 2.4B active | 22 scheduled Indian languages and English | Apache 2.0 | Trained from scratch in India on IndiaAI compute (Sarvam's account) |
Sarvam 105B | Feb 2026; weights 6 Mar 2026 | 106B total, 10.3B active; 128K context | As Sarvam 30B | Apache 2.0 | As Sarvam 30B |
Prism (Sarvam AI) | 3 Nov 2026 | Built on Sarvam 105B | As Sarvam 105B | Apache 2.0 | “Decision” model: picks from options you supply; writes no text |
Sarvam-105B-1711 | 15 Nov 2026 | As Sarvam 105B | As Sarvam 105B | Apache 2.0 | More training for tool use, agents and Indian languages; FP4 and BF16 builds |
Param2 17B (BharatGen) | Feb 2026 | 17B total, 2.4B active; 32K context | English, Hindi and 21 more Indian languages | BharatGen non-commercial | Built from scratch in India (BharatGen's account) |
Sarvam weighted training towards the 10 most-spoken languages, so test yours.
What GPU Sarvam and Param2 need
Three rules size any model (our GPU memory guide has the maths):
- Fit. The weights (a model's learned numbers) must fit in GPU memory: parameters times bytes per parameter. Full precision (BF16) uses 2 bytes, FP8 1 and FP4 about half a byte; lower precision shrinks the model with a small loss of quality.
- Headroom. Our rule of thumb: keep the weights to about 80% of GPU memory, leaving the rest for the KV cache (the model's working memory per conversation); long contexts and many users need more.
- MoE. Total size sets the memory and active size sets the speed. All 106B parameters of Sarvam 105B must sit in memory, though only 10.3B work on each token.
Model and precision | Weights | Smallest GPU count that holds it (sizing only) | Production setup |
|---|---|---|---|
Sarvam 105B, BF16 | About 212 GB | 4 x H100 80 GB (Sarvam's example splits it across four GPUs) | |
Sarvam 105B, FP8 | About 106 GB | 1 x H200 (141 GB) or B200 (180 GB) | 8-GPU H200 or B200 server; B300 (2.3 TB per server) for many users or long chats |
Sarvam-105B-1711, FP4 | About 70 GB (download size) | 1 x B200 or B300 (of the GPUs we source, only B200 and B300 run FP4 natively) | 8-GPU B200 or B300 server |
Prism, FP8 | About 106 GB | 1 x H200 or B200 (needs about 120 GB free, says Sarvam) | 8-GPU H200 or B200 server; Sarvam reports a 44 ms median reply on one B200 (vendor-reported) |
Sarvam 30B, BF16 | About 64 GB | 1 x H200 | Several copies on one 8-GPU server |
Sarvam 30B, FP8 | About 32 GB | 1 x H100 | Several copies on one 8-GPU server |
Param2 17B, BF16 | About 34 GB | 1 x H100 | Shares an 8-GPU server with other models |
These are our estimates, for sizing only: we source whole 8-GPU servers, and GPU counts on the model cards are examples, not minimums. One 8-GPU server can run several copies of Sarvam 30B or 105B, or an Indian model alongside a global one. Fine-tuning (further training on your own data) with LoRA, which trains small add-on layers instead of the whole model, fits on one 8-GPU server for Sarvam 30B or 105B. A full fine-tune needs roughly 16 bytes per parameter, about 1.7 TB for Sarvam 105B plus working memory, so plan a cluster of several 8-GPU servers or one B300 server with careful setup. See multi-server B300 clusters and B300 fine-tuning costs, or compare the B300 and B200 for many users.
How good are Indian models?
Built for Indian languages, still behind global open models on independent tests. Artificial Analysis, an independent benchmarking firm, scored Sarvam's models on its Intelligence Index (mostly English tasks, reasoning mode):
Artificial Analysis Intelligence Index v4 (published 2 Apr 2026) | Maker | Score |
|---|---|---|
gpt-oss-120B (high reasoning setting) | OpenAI, US | 33 |
GLM-4.5-Air | Zhipu AI (Z.ai), China | 23 |
Qwen3 30B A3B 2507 | Alibaba, China | 22 |
Sarvam 105B | Sarvam AI, India | 18 |
Sarvam 30B | Sarvam AI, India | 12 |
Sarvam-M | Sarvam AI, India | 8 |
These are version 4 scores. Artificial Analysis moved to version 5 on 27 October 2026 and shows no v5 score for Sarvam 105B, so do not compare them with v5 scores, including those elsewhere in this series. Sarvam's own tests (vendor-reported) tell a similar story:
Test (what it measures) | Sarvam 105B | Comparison in Sarvam's table | Reading |
|---|---|---|---|
MMLU (general knowledge, English) | 90.6 | gpt-oss-120B: 90.0 | Roughly level |
SWE-Bench Verified (fixing real code) | 45.0 | gpt-oss-120B: 50.6; GLM-4.5-Air: 57.6 | Behind |
Arena Hard v2 (hard chat prompts) | 71.0 | gpt-oss-120B: 88.5 | Behind |
Indian-language win rate (110 prompts in 22 languages, Gemini 3 as judge) | 90% average (Sarvam 30B: 89%) | Opposing models not named | Strong claim, not independently checked |
We found no independent leaderboard for Indian-language text yet. Early software gaps are closed: vLLM, the popular serving software, has supported Sarvam 30B and 105B natively since 7 March 2026, and official compressed builds (FP8 for servers, GGUF for small machines) followed later that month.
Our advice:
- Indian-language chat, search and voice front ends: an Indian model.
- Code and multi-step reasoning: a global open model.
- Before you choose: test 100 of your own prompts in your own languages.
Why Indian languages cost more per word
APIs charge per token, and a GPU produces a fixed number of tokens a second (our token pricing guide explains the billing). Each model's tokenizer decides how text is cut into tokens, and many cut Indian scripts into far more pieces than English. A July 2026 preprint, "The Tokenizer Tax", counted tokens per word on the same professionally translated sentences (FLORES-200):
Tokens per word vs English (English = 1.00) | GPT-4 tokenizer | GPT-4o tokenizer | Qwen2.5 tokenizer |
|---|---|---|---|
Hindi | 4.08 | 1.34 | 3.79 |
Marathi | 5.84 | 2.11 | 5.33 |
Bengali | 6.52 | 1.91 | 5.65 |
Gujarati | 8.00 | 1.87 | 7.03 |
Tamil | 9.89 | 2.57 | 7.92 |
Telugu | 10.68 | 2.49 | 9.02 |
Kannada | 12.06 | 2.72 | 9.42 |
Malayalam | 13.04 | 2.85 | 10.54 |
Per-word ratios are a different measure from the totals in our token pricing guide, which count every token in whole sentences. On that basis, GPT-4's tokenizer needed about 4.8 times as many tokens as English for Hindi and 9 times for Malayalam; use totals like these, not per-word ratios, to estimate cost. The gap is a design choice: across Indian languages, GPT-4o's tokenizer needs on average about 73% fewer tokens per word than GPT-4's. Sarvam says the tokenizer it built for Sarvam-1 in 2024 needs 1.4 to 2.1 tokens per word in 10 Indian languages, against about 1.4 for English (vendor-reported).
On an API, more tokens means a bigger bill; on your own server, fewer answers per GPU-hour. Before you budget, count tokens on your own prompts with each candidate model's tokenizer.
IndiaAI GPU rates and who can apply
The government buys GPU time from empanelled providers and offers it to eligible Indian users at subsidised rates through the IndiaAI compute portal. If you qualify, apply.
Item | Figure | Source and date |
|---|---|---|
Approved | ₹10,371.92 crore (a crore is 10 million rupees) over five years; a target of 10,000 or more GPUs through public-private partnership | Union Cabinet, 7 Mar 2024 |
Portal opened | Up to 40% subsidy for eligible users; about 14,000 GPUs live at launch | MeitY launch, 6 Mar 2025 |
GPUs | 38,231 onboarded; IndiaAI's empanelment list shows 44,077 accelerators offered by 13 bidders (offered, not necessarily live) | Lok Sabha answer, 25 Mar 2026; IndiaAI list dated 3 Nov 2026 |
Home-grown models backed | 20 proposals: 12 large models and 8 small ones | Parliament answers, Jul and Aug 2026 |
Where the subsidy has gone | ₹1,058 crore of ₹1,275 crore committed (about 83%) to nine foundation-model builders; ₹10.14 crore (0.8%) to researchers, students and fellows | ThePrint, portal data to 8 Oct 2026 |
That split follows the design: building India's own foundation models is a main aim, and the government has paid about 99% of those builders' bills so far (ThePrint). Sarvam, for one, was allocated 4,096 H100s for six months in 2025, and its ₹246.72 crore bill is fully subsidised.
Who can apply: startups and MSMEs working in AI that have DPIIT recognition, ₹50 lakh of annual revenue or ₹1 crore of funding; researchers meeting citation thresholds; students, IndiaAI fellows and government bodies. Requests go in from the 1st to the 25th of each month; those under 5,000 GPU-hours are approved automatically, and larger ones go to a committee.
Reference on-demand rates (13 Mar 2026), eligible users only | Per hour, whole 8-GPU server | Per GPU-hour |
|---|---|---|
H100 SXM | ₹1,224 | ₹153 (about US$1.61) |
H200 SXM | ₹1,125 | ₹140.63 (about US$1.48) |
B200 SXM | ₹2,584 | ₹323 (about US$3.39) |
A100 80GB | ₹1,087.20 | ₹135.90 (about US$1.43) |
IndiaAI calls these indicative and does not say whether they are before or after the subsidy; they are not market prices. GPU businesses can read how they shape the market in our guide for GPU clouds in India.
API, IndiaAI or your own GPUs?
The break-even between paying per token and running your own server is in our LLM sizing guide and rent, buy or API guide.
Route | Good fit when | Watch out for |
|---|---|---|
Sarvam's own API (prices below) | Low or spiky volume; quick pilots | Introductory prices end with 2026; prompts go to a third party; Sarvam 30B is marked for retirement there |
IndiaAI compute portal | You are an eligible Indian startup, researcher or public body | Monthly approval cycle; large requests go to a committee |
Your own dedicated 8-GPU server in India | Steady volume; fine-tuning on company data; data that must stay in India; real-time voice | Whole 8-GPU servers, terms of a year or more, billed upfront; costs the same busy or idle; you run the serving software (vLLM or SGLang) |
Sarvam's API prices, before GST: Sarvam 105B costs ₹10 per million input tokens (₹0.25 cached) and ₹25 per million output tokens; Prism costs ₹4 per million input tokens, with no output charge. OpenRouter, a marketplace for many models' APIs, listed no model from an Indian lab when we checked on 19 November 2026.
Why run Indian models in India
India has no blanket rule that data must stay in the country, but sector rules point regulated buyers there. Payment system data must be stored only in India (Reserve Bank of India, since 2018). Government data on cloud must stay in India, with providers empanelled by MeitY, the IT ministry. Under India's privacy law, the DPDP Act, the government can require some large data handlers to keep specified data in India; the main duties apply from 13 May 2027. For a bank, insurer or government supplier, an open model on a server in India that only you use is the easiest setup to explain to an auditor (see bare metal vs cloud GPUs). General information, not legal advice: check with counsel, and see hosting your GPUs in India.
Distance matters for voice: the big Indian cities are 13 to 27 ms from Mumbai on the public internet.
Running Sarvam models on GPUs in India
Sarvam and other Indian-language models whose data has to stay in the country need servers in India, one of our six regions; FlexiCloud itself is based in Kochi. We source single-tenant 8-GPU H100, H200, B200 and B300 servers there from verified operators, on terms of a year or more paid upfront, not hourly, sized to your model, context and traffic. Your team runs the serving software. If anyone we recommend pays us, we tell you in writing before you sign. See B300 servers in India and our India data centre guide.
For a quote in India, send us your requirements or book a call.
Where these figures come from
Checked on 19 November 2026 unless a date is given. US dollars at ₹95.25 per US$, as on our other India pages.
- Sarvam: 30B and 105B launch post (6 Mar 2026; benchmarks vendor-reported), Hugging Face cards for Sarvam 105B, 30B, Sarvam-1, Prism and 105B-1711; posts on Sarvam-M, Prism and the Sarvam-1 tokenizer; API pricing, Sarvam 30B deprecation notice.
- Param2: BharatGen's Hugging Face card.
- Scores: Artificial Analysis (2 Apr 2026) and its v5 note (27 Oct 2026). vLLM support.
- Tokens: The Tokenizer Tax (27 Jul 2026, Table 2); Petrov et al. (2023, ratios our arithmetic); Microsoft (20 May 2024).
- IndiaAI: Ready Reckoner (rates dated 13 Mar 2026), eligibility, empanelled GPUs (list dated 3 Nov 2026), allocations; reported by TelecomTalk (Cabinet approval, 7 Mar 2024), Inc42 (portal launch and subsidy, 6 Mar 2025), IANS via Morung Express (14,000 GPUs live at launch, 6 Mar 2025), IANS (Lok Sabha, 25 Mar 2026), Asianet Newsable (Parliament, 5 Aug 2026) and ThePrint (20 Oct 2026).
- Data rules: RBI (2018); MeitY cloud rules, reported by Digit (2017); DPDP Rules (PIB, Nov 2025), timeline explained by AZB & Partners (Nov 2025).
- Latency: public pings, in our India hosting guide. OpenRouter model list.