Sarvam on Friday, March 6, introduced the rollout of its two foundational multi-lingual AI fashions that have been unveiled on the just lately held India-AI Influence Summit 2026, beneath an open-source licence.
The 30 billion- and 105 billion-parameter massive language fashions (LLMs) are reasoning fashions which have been constructed from scratch by coaching them on large-scale, high-quality datasets curated in-house, the Indian AI startup mentioned in a weblog submit. Each fashions have been skilled utilizing compute from GPUs (Graphics Processing Models) made obtainable beneath the Indian government-backed Rs 10,372-crore IndiaAI Mission with infrastructure assist from knowledge heart operator Yotta and technical assist from Nvidia, Sarvam mentioned.
Whereas the 2 AI fashions have been first launched on the AI Influence Summit 2026 hosted by India in New Delhi final month, Sarvam has now made these fashions obtainable for industrial use beneath the Apache 2.0 open-source licence, with the mannequin weights obtainable for obtain on AIKosh and Hugging Face platforms. Each fashions are additionally accessible by way of Sarvam’s Indus AI chatbot app and thru the corporate’s API developer dashboard.
In current days, Sarvam has emerged because the flagbearer of India’s ‘sovereign AI’ push, because the central authorities seeks to cut back dependence on overseas AI giants equivalent to OpenAI and Anthropic by enabling the event of smaller, environment friendly fashions which can be tailor-made to native Indian languages and use circumstances.
Nonetheless, some observers have additionally questioned whether or not so-called sovereign AI fashions might be open-weight, as permitting anybody on the earth to freely modify and distribute them raises a key query about what precisely constitutes sovereignty within the context of AI.
Internally, Sarvam mentioned that the 30B mannequin is used to energy its conversational agent platform referred to as Samvaad whereas the bigger, 105B mannequin is the inspiration for its Indus AI assistant constructed for advanced reasoning and agentic workflows. The 2 fashions have additionally been optimised to be deployed throughout a variety of {hardware}, together with private gadgets like laptops.
“Constructing these fashions required creating end-to-end functionality throughout knowledge, coaching, inference, and product deployment. With that basis in place, we’re able to scale to considerably bigger and extra succesful fashions, together with fashions specialised for coding, agentic, and multimodal conversational duties,” Sarvam wrote within the weblog submit.
Story continues beneath this advert
Underneath the hood
The 30-billion- and 105-billion-parameter fashions use a mixture-of-experts (MoE) transformer structure, which prompts solely a fraction of their complete parameters at a time, considerably lowering computing prices, Sarvam mentioned. The 30B mannequin helps a 32,000-token context window aimed toward real-time conversational use, whereas the bigger mannequin provides a 128,000-token window for extra advanced, multi-step reasoning duties.
When it comes to effectivity, Sarvam 30B makes use of Grouped Question Consideration (GQA) to cut back KV-cache reminiscence whereas sustaining robust efficiency. Sarvam 105B, however, depends on DeepSeek-style Multi-head Latent Consideration (MLA) that additional reduces reminiscence necessities for long-context inference.
The information used to coach each fashions consists of code, normal internet knowledge, specialised data corpora, arithmetic, and multilingual content material. Sarvam mentioned {that a} substantial portion of the coaching finances was allotted towards curating a multilingual corpus of knowledge within the 10 most-spoken Indian languages.
Efficiency on benchmarks
The Sarvam 105B mannequin carried out higher than the 30B mannequin on benchmarks in the course of the early phases of coaching, which suggests environment friendly scaling behaviour, Sarvam mentioned.
Story continues beneath this advert
When in comparison with LLMs of comparable dimension, the 105B mannequin achieved outcomes much like pt-oss 120B and Qwen3-Subsequent (80B) on normal capabilities. It additionally demonstrates robust efficiency on agentic reasoning and job completion, outperforming DeepSeek R1, Gemini 2.5 Flash, and o4-mini on Tau 2 Bench.
Nonetheless, Sarvam 105B might not be the strongest code-generation mannequin as its efficiency on SWE-Bench Verified lagged behind in contrast fashions. As for the smaller 30B mannequin, the outcomes confirmed that when in comparison with Nemotron 3 Nano 30B, Sarvam’s AI mannequin is barely forward in coding (SWE-Bench Verified) and agentic reasoning (Tau2) however barely worse in different benchmarks equivalent to Stay Code Bench v6 and BrowseComp.
Apparently, Sarvam mentioned that its 30B mannequin will get 20 to 40 per cent extra tokens/sec throughput in comparison with Qwen3 resulting from code and kernel optimisations. Sarvam’s efficiency on Indian languages is aided by its tokenizer that was constructed and skilled from scratch for environment friendly tokenization throughout all 22 scheduled Indian languages, spanning 12 completely different scripts.
Based mostly on the fertility rating, which is the typical variety of tokens required to characterize a phrase, Sarvam’s tokenizer outperformed different open-source tokenizers in encoding Indic textual content effectively.
Story continues beneath this advert
Security and safety
Within the supervised fine-tuning stage, Sarvam mentioned that it fine-tuned each fashions on a dataset masking commonplace and India-specific danger eventualities. The dataset additionally included adversarial and jailbreak-style prompts mined by automated red-teaming. These prompts have been paired with policy-aligned, protected completions for supervised coaching, as per the corporate.