Sarvam-105B: India's Open-Source Reasoning Model Guide

What Is Sarvam-105B

Sarvam-105B is a 105-billion parameter open-source reasoning model developed by Sarvam AI, a Bangalore-based artificial intelligence company founded with the explicit mission of building India-first AI infrastructure. The model was trained entirely in India using domestic GPU clusters, making it the first frontier-class reasoning model to emerge from the Indian AI ecosystem.

The model supports a 128K token context window, enabling it to process long documents, codebases, and multi-turn conversations without truncation. It uses a decoder-only transformer architecture with mixture-of-experts (MoE) design, allowing efficient inference by activating only a subset of parameters for each token. This architectural choice means that while the total parameter count is 105 billion, the active parameters during inference are substantially lower, reducing compute costs per query.

Sarvam AI was founded by Vivek Raghavan and Pratyush Kumar, both with deep roots in Indian AI research. The company raised significant funding specifically to build India-first AI infrastructure, rejecting the common approach of fine-tuning Western models for Indian languages. Their thesis is that genuine multilingual AI requires training from scratch with Indian language data woven into the pretraining corpus, not bolted on afterward.

Architecture and Training Infrastructure

The technical architecture of Sarvam-105B reflects deliberate decisions about how to build a competitive reasoning model within the constraints of Indian GPU infrastructure. The model uses a sparse mixture-of-experts (MoE) transformer architecture, where each token is routed to a subset of expert networks rather than passing through all parameters. This design enables the model to have 105 billion total parameters while keeping per-token compute costs comparable to a dense model roughly one-third the size.

Training was conducted entirely on GPU clusters located in India, a decision driven by both sovereignty goals and practical considerations around data residency. The training data included a carefully curated mixture of English text, Indian language corpora, code, and reasoning-focused datasets. Sarvam AI invested heavily in data collection and curation for Indian languages, building custom pipelines to process text from government documents, academic papers, news outlets, and web crawls across more than 22 languages.

The training process used a multi-stage approach. The initial pretraining phase focused on general language understanding across English and Indian languages. Subsequent stages progressively introduced reasoning-specific data including mathematical problem sets, code generation tasks, logical deduction exercises, and chain-of-thought reasoning traces. This staged approach follows patterns established by models like DeepSeek V3 but adapts them for a multilingual context where reasoning capability needs to transfer across language boundaries.

A key innovation in Sarvam's training pipeline is cross-lingual reasoning alignment. The model was explicitly trained to maintain reasoning quality when switching between languages mid-conversation or when the reasoning task is presented in one language and the response is expected in another. This is a harder problem than monolingual reasoning and required specific data curation strategies that Western labs have generally not prioritized.

Sarvam-30B: The Efficient Alternative

Alongside the flagship 105B model, Sarvam AI released Sarvam-30B, a 30-billion parameter dense model that retains approximately 85% of the larger model's reasoning capability. While Sarvam-105B demonstrates what Indian AI infrastructure can produce at frontier scale, Sarvam-30B addresses the practical question of deployment cost and accessibility.

The 30B model uses a dense transformer architecture rather than the MoE design of the 105B variant. This means all 30 billion parameters are active during inference, but the smaller total size makes it deployable on a single A100 GPU or a pair of A10G GPUs. For Indian startups and enterprises that cannot justify the cost of multi-GPU clusters, Sarvam-30B makes frontier-adjacent multilingual reasoning accessible at a fraction of the infrastructure cost.

The efficiency gains are substantial. At the 30B parameter scale, inference latency drops significantly compared to the 105B model, making real-time applications like chatbots, code assistants, and customer support systems practical. The model also supports quantization down to 4-bit precision with minimal quality degradation, further reducing deployment requirements. For teams comparing efficient model options, the 30B variant sits in a similar deployment class to models like Qwen 3.5 models but with dramatically better Indian language performance.

Reasoning Benchmarks and Performance

Sarvam-105B was evaluated against standard AI benchmarks and shows competitive performance with established frontier models. The benchmark results position it as a credible alternative for organizations that need strong reasoning combined with Indian language support, though it does not uniformly surpass the leading Western and Chinese models on every English-centric metric.

The most significant benchmark results are in the multilingual categories. On IndicBench and similar Indian language evaluation suites, Sarvam-105B substantially outperforms models like GPT-4, Claude, and Gemini that were primarily trained on English-centric data. This is not surprising given the training data composition, but it validates the thesis that building for Indian languages from the ground up produces fundamentally better results than fine-tuning English-first models.

On English-centric benchmarks like MMLU, GSM8K, and HumanEval, Sarvam-105B performs competitively with models in the 70-100B parameter class. It does not claim to beat GPT-4o or Claude Sonnet on pure English reasoning, and Sarvam AI has been transparent about this positioning. The value proposition is not “better than Western models at everything” but rather “competitive on English tasks while being dramatically better for Indian language workloads.”

Multilingual Indian Language Support

India has 22 officially recognized languages and hundreds of additional dialects spoken by over 1.4 billion people. Most frontier AI models treat Indian languages as an afterthought, adding support through translation layers or limited fine-tuning that produces stilted, unnatural output. Sarvam-105B takes a fundamentally different approach by incorporating Indian language data directly into pretraining.

The practical difference between Sarvam's native multilingual approach and the fine-tuning approach used by Western models is stark. When asked to explain a complex technical concept in Tamil, a fine-tuned English model typically translates English reasoning into Tamil, producing grammatically correct but unnatural output. Sarvam-105B reasons directly in Tamil, producing explanations that read as if written by a fluent Tamil speaker. This distinction matters enormously for production applications serving Indian users.

Code-switching support deserves special attention. In everyday Indian communication, people routinely mix Hindi and English (Hinglish), Tamil and English (Tanglish), and other language combinations. Western models generally struggle with this because their training data does not include sufficient code-switched text. Sarvam-105B handles these mixed-language inputs fluently because the pretraining corpus was specifically curated to include authentic code-switched text from social media, messaging platforms, and web forums. The ability to process real-world Indian communication patterns, not just formal single-language text, is what makes these models practically useful for Indian market applications.

Indus AI Assistant and Deployment

Sarvam AI does not just release model weights and walk away. The company also ships Indus, a consumer-facing AI assistant powered by the Sarvam model family. Indus serves as both a product and a live demonstration of what these models can do in production, handling conversational AI tasks in Indian languages including information retrieval, content generation, code assistance, summarization, and general reasoning.

For developers and enterprises looking to deploy Sarvam models on their own infrastructure, both models are available on Hugging Face with full model weights and tokenizer files. They are compatible with standard inference frameworks including vLLM, Text Generation Inference (TGI), and llama.cpp for quantized deployments. Sarvam AI also provides an API for teams that prefer managed inference without operating their own GPU infrastructure.

The deployment flexibility is a direct consequence of the Apache 2.0 license. Unlike restricted-license models that prohibit commercial use above certain user thresholds or require revenue sharing, Sarvam's models can be deployed without any legal constraints. This is particularly important for Indian government agencies, financial institutions, and healthcare organizations that need on-premises deployment for regulatory compliance. The open-source model landscape continues to expand with similar permissive licensing, as seen with models like GLM-5 from Zhipu AI, which follows the same Apache 2.0 approach for the Chinese market.

Indian AI Sovereignty Implications

Sarvam-105B is significant beyond its technical merits. It represents the first time a frontier-class reasoning model has been developed entirely within India, using Indian infrastructure, Indian talent, and data pipelines designed for Indian languages. This matters for several interconnected reasons that extend well beyond benchmark scores.

The Indian government has been increasingly vocal about AI sovereignty, with the Ministry of Electronics and Information Technology actively supporting domestic AI development through funding programs and compute infrastructure investments. Sarvam-105B aligns directly with these national priorities. It provides a concrete example of what Indian-built AI infrastructure can achieve, strengthening the case for continued investment in domestic capabilities.

The open-source release strategy amplifies the sovereignty impact. By releasing under Apache 2.0, Sarvam AI ensures that the models cannot be restricted or revoked. Any Indian organization can download, deploy, and build on these models permanently, regardless of what happens to Sarvam AI as a company or how geopolitical relationships evolve. This is infrastructure that, once released, cannot be taken back. The pattern mirrors the broader shift toward open-source frontier models globally, where organizations like DeepSeek and Mistral have demonstrated that open weights can coexist with commercial viability.

The competitive dynamics are also worth noting. India is the world's largest democracy and its fifth-largest economy. Having domestic frontier AI capability positions India as a third pole in global AI development alongside the US and China. For international businesses operating in India, Sarvam's models offer a compliance-friendly option that keeps data within Indian jurisdiction while providing reasoning capability comparable to foreign alternatives. This is not just a technical achievement but a strategic one that reshapes how the global AI map looks.