Skip to content
BharatGen's Param2, Released in February 2026, Is a 17-Billion-Parameter Model for English and All 22 Scheduled Indian Languages That Uses 2.4 Billion Parameters for Each Token
AI & Technology

BharatGen's Param2, Released in February 2026, Is a 17-Billion-Parameter Model for English and All 22 Scheduled Indian Languages That Uses 2.4 Billion Parameters for Each Token

Illustration by tuput

English

BharatGen, the consortium anchored at IIT Bombay, released Param2 at the India AI Impact Summit in February 2026. It is a mixture-of-experts reasoning model that BharatGen says it pretrained from scratch on about 22.5 trillion tokens, with weights posted for download under a non-commercial research licence.

· · 9 min read

BharatGen, the government-funded consortium anchored at IIT Bombay, released Param2 at the India AI Impact Summit in New Delhi in the week of 16 February 2026. It is a language model with 17 billion parameters, the adjustable numbers a model learns in training, of which about 2.4 billion work on any one token, a word or a piece of a word. The consortium says it pretrained the model from scratch, not from someone else’s model, and that it handles English and all 22 languages in India’s Eighth Schedule. Weights are on Hugging Face and on AIKosh, the consortium says; AIKosh is the IndiaAI Mission’s library of models and datasets.

Sources differ on the day. Business Today reports the launch on 16 February, BharatGen’s launch post and the Press Information Bureau (PIB) release are dated 17 February, and Business Standard reported the announcement on 18 February. Every performance figure below is BharatGen’s own, and none of the pages consulted reports a test of Param2 run by anyone outside the consortium.

Nine institutions, two sources of money

Nine institutions make up the consortium, according to its two-year retrospective of 15 October 2026: IIT Bombay, IIIT Hyderabad, IIT Mandi, IIT Kanpur, IIT Hyderabad, IIM Indore, IIIT Delhi, IIT Madras and IIT Kharagpur. The Department of Science and Technology has backed it since its launch in New Delhi in October 2024. PIB’s release of 17 February 2026 gives the funding as Rs 235 crore from the department’s National Mission on Interdisciplinary Cyber-Physical Systems and a further outlay of Rs 1,058 crore from the IndiaAI Mission. How that mission’s grants work is set out in a piece on the IndiaAI Mission’s GPUs, models and money.

Param2 follows Param1, a 2.9-billion-parameter model for Hindi and English that BharatGen described in an arXiv preprint on 16 July 2025. Param1 trained on 7.5 trillion tokens, the retrospective says. Seven months on, Param2 has nearly six times the total parameters, about three times the training text and 23 languages in place of two, by tuput’s arithmetic from those figures.

The compute came through the IndiaAI Mission, by BharatGen’s account. Rishi Bal, its chief executive, told Business Today on 16 February that the GPUs used so far from the mission had gone into training Param2 and that the project had used roughly 4,000 of them for over a month. Training took 24 days on NVIDIA H100 GPUs using NVIDIA’s NeMo software, says the model card, and Business Standard reported that NVIDIA collaborated on the training pipeline. Neither the card nor the post gives a GPU count, and the 576 hours the card lists as GPU hours equal 24 days of elapsed time, so the two accounts cannot be compared from the published figures.

Seventeen billion stored, 2.4 billion at work

Architecturally it is a mixture-of-experts model. Instead of passing every token through the whole network, a small router picks a few specialist sub-networks, called experts, for each one. Hugging Face’s configuration file lists 64 routed experts with 6 chosen per token, plus 2 shared experts that always run, across 21 layers, the first of them an ordinary dense layer. Attention uses 32 query heads that share 8 key and value heads. A mixture-of-experts explainer follows the design from a 1991 paper to Sarvam’s 105-billion-parameter model.

A hospital is BharatGen’s own analogy: a patient sees a general physician and a few relevant specialists, not every doctor in the building. Per the model card, the two shared experts keep a stable representation across languages, which helps with code-switching, the mixing of Hindi and English in one sentence, and lets what the model learns in one language carry over to others. Compute goes to only about 2.4 billion parameters per token, BharatGen says, so the model runs at the cost of a much smaller one.

Memory is the cost. Every parameter stays loaded even though a fraction work at a time, and the Hugging Face page counts 17,151,125,376 parameters in the stored weights, in files totalling 34.3 GB, which matches the 34 GB of GPU memory BharatGen’s post gives for loading the model in bfloat16, a 16-bit number format.

What it was trained on

Three pretraining phases appear on the model card. Phase one covers 14.05 trillion tokens, with English at 8.5 trillion, code at 1.83 trillion, 998.6 billion Hindi tokens and 1.5 trillion tokens in the other Indian languages combined. Hindi and the other Indian languages make up about 18 per cent of it, by tuput’s arithmetic. Phase two leans on English, code, mathematics, instruction and reasoning data, and a 500-billion-token third phase is labelled long-context pretraining. Indian-language data was filtered with quality classifiers built for 22 languages, BharatGen says.

About 22.5 trillion tokens is the card’s headline figure, and its combined total reads 22.63 trillion. The phase two rows it lists add up to 7.14 trillion tokens against a stated phase total of 8.08 trillion, a gap of about 0.94 trillion that the card does not explain. Analytics India Magazine’s report of the launch gives 22 trillion.

A tokenizer is the piece that cuts text into tokens, and Param2 has its own. Its vocabulary holds 128,000 entries, BharatGen says, built by byte pair encoding and trained on all the target languages, so that Indian-language words split into fewer pieces than in global models. No pieces-per-word figures appear in the post. What other teams have measured is covered in a piece on tokenizer fertility for Indian languages. Its context window, the amount of text it can read at once, is 32,768 tokens.

A reasoning model with Telugu comments in its code

The version posted on Hugging Face is a reasoning model. Before its answer it writes a step-by-step trace between think tags, and the repository includes a small parser that separates the trace, any tool calls and the final answer. Code, mathematics and tool calling are also covered, the post says, and the model card shows three sample outputs: a multi-step average problem, a Python binary search with every comment written in Telugu, and an answer in Hindi on why languages, food and festivals differ across Indian states.

The 22 languages are those of the Eighth Schedule, which our explainer lists. Hindi, Marathi, Gujarati and Tamil are among the strongest, the post says, and quality varies with how much training data a language has.

The scores, and two tables that differ

Four open models serve as comparisons in both: Qwen1.5-MoE-A2.7B, DeepSeek-R1-Distill-Qwen-14B, DeepSeek-V2-Lite and gpt-oss-20b. Scores are zero-shot, meaning the prompt carries no worked examples, run with the LM Evaluation Harness, an open testing toolkit, and with thinking switched off for every model. That means the numbers do not capture what the reasoning trace adds, the post says.

On English, maths and code the post and the card agree:

TestParam2 scoreRank among the five models
ARC Challenge (English)56.831st
HellaSwag (English)77.432nd
MMLU (English)57.794th
GSM8K (maths)57.323rd
HumanEval (code)36.591st
IFEval (instruction following)34.052nd

On Indian-language tests the two differ. Param2 comes first on five in the launch post of 17 February, with MMLU in Hindi at 59.23, HellaSwag in Hindi at 49.63, Indic BoolQ at 75.98, ARC Challenge (Indic) at 36.37 and TriviaQA (Indic multiple choice) at 72.95, and third on the Sanskriti cultural-knowledge test at 66.54. The model card as read on 19 November 2026 gives 64.1, 64.1, 87.7, 90.1, 84.14 and 88.6 for the same six tests in that order, first on all six. Thinking was off in both, and the two Indic averages cover 10 languages in each version. Hugging Face shows the card last modified on 18 November 2026, and it does not say whether the model or the test settings changed. The text-models page still carries the post’s figures, which it labels the February release.

The one outside document that discusses Param2’s published results is a survey, not a test run. In a preprint of 12 August 2026 (revised 16 August), Avinash Agarwal and Vridhi Jain compared Indian foundation models using only publicly reported scores. The authors write that Param2 reports limited results, concentrated on Indic-language tests, and that most results reported for Indian models are the developers’ own. Param2’s training is classed as fully indigenous on a provisional basis, and the authors say they did not verify every element of that class independently.

The technical reports page lists papers on Param1, AyurParam and a farm-advice assistant, and a September 2026 white paper on a sovereign AI stack that evaluated Param2 for reasoning. None is on Param2 itself as of 19 November 2026. The post calls the release an early post-training checkpoint that needs further fine-tuning and safety alignment before production use, and Bal told Business Today it is a checkpoint developers can fine-tune.

Who can download it, and on what terms

The Hugging Face repository was created on 16 February 2026 and last modified on 18 November 2026. The page showed 167,875 downloads in all and 86 likes on 19 November. Its own text-models page says 167,000 or more downloads for Param2 and 219,000 or more for the Param family. BharatGen lists hospitals, state governments, a ministry and a newsroom as partners building on its models, and its pages do not say which of them use Param2. The tuput piece on EKA-1 and Dhīr covers two other model releases from the IndiaAI Mission.

BharatGen’s pages call the release open weights. Its licence is the BharatGen Research License, last updated on 7 August 2025, and it allows use, modification and distribution for non-commercial research and academic purposes. It requires written permission from BharatGen’s legal team for commercial use, for publicly releasing, fine-tuning or deploying the model or its derivatives, and for offering it through any hosted service or API, even a free one. Users own the outputs they generate, subject to the non-commercial limit. A company that wants to build a product on Param2 therefore writes to BharatGen first.

Bal said in February to expect “progressively larger models over the course of this year”. As of 19 November 2026 BharatGen’s text-models page still lists Param2 as its flagship, and the model card still says a variant with a 128k-token context window “will be released soon”.

Share
Copied!

Sources & further reading

  1. BharatGen: Param2, what we built and how (17 February 2026)
  2. Hugging Face: bharatgenai/Param2-17B-A2.4B-Thinking, model page and card (read 19 November 2026)
  3. Hugging Face: Param2-17B-A2.4B-Thinking, config.json
  4. Hugging Face: Param2-17B-A2.4B-Thinking, LICENSE (BharatGen Research License, 7 August 2025)
  5. PIB: The Artificial Intelligence is no longer an option but an essential component of working in every sphere of life, Dr Jitendra Singh at the BharatGen session (17 February 2026)
  6. BharatGen: Param, language models for India (text models page, read 19 November 2026)
  7. BharatGen: Two years of BharatGen (15 October 2026)
  8. BharatGen: Technical reports (read 19 November 2026)
  9. arXiv 2507.13390: PARAM-1 BharatGen 2.9B Model (16 July 2025)
  10. arXiv 2608.11891: Benchmark-Based Comparative Assessment of Publicly Benchmarked Indian Foundation Models, Agarwal and Jain (12 August 2026, revised 16 August)
  11. Business Today: BharatGen to launch 17-billion-parameter multilingual AI model at India AI Impact Summit (12 February 2026)
  12. Business Today: Most real-world use cases don't need trillion-parameter AI, BharatGen CEO Rishi Bal on Param 2 launch (16 February 2026)
  13. Analytics India Magazine: After Param2, BharatGen unveils Patram, Sooktam and Shrutam AI models at India AI Impact Summit (17 February 2026)
  14. Business Standard: BharatGen unveils Param2 17B multilingual MoE model under sovereign AI push (18 February 2026)

Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.

#param2#bharatgen#indian language models#mixture of experts#iit bombay#india ai impact summit#indiaai mission#hugging face

Enjoyed this? Get the next one.

One good read at a time, straight to your inbox. No spam, unsubscribe anytime.

More in AI & Technology
The IndiaAI Mission Has Over 38,000 GPUs at Rs 65 an Hour, 12 Funded Model Builders and Open Models From Sarvam Trained on Mission Compute, With 20,000 More GPUs Announced in February 2026
Compute is priced at Rs 65 an hour, the first models are out and the February summit drew more than 100 countries. What the government's own papers say about each pillar of the mission, and the views of Carnegie India, Takshashila and an IISc-linked research head.
Sarvam's 105-Billion-Parameter Model Computes With 10.3 Billion of Them for Each Token, Using the Mixture-of-Experts Design That Runs From a 1991 Paper to Mixtral and DeepSeek
A small router sends every word to a few of 128 sub-networks, so most of the model sits idle for any one word. Here is where the design came from, how Mixtral, DeepSeek and Sarvam use it, and what it does not save.
DeepSeek's V4.1-Flash, an Open-Weight 552-Billion-Parameter Model Released on 10 September 2026, Cuts Its Per-Token Memory Cache to 890 Bytes, About a Quarter of Its Predecessor's
A model's short-term memory of a long conversation is expensive to keep. DeepSeek's September release compresses it by design, and its own report says where the model still trails. Here is the mechanism, the timeline from 2024, and what outsiders have checked.
← all articles