Open-Sourcing Sarvam 30B and 105B

We’re releasing Sarvam 30B and Sarvam 105B as open-source models. Both are reasoning models trained from scratch on large-scale, high-quality datasets curated in-house across every stage of training: pre-training, supervised fine-tuning, and reinforcement learning. Training was conducted entirely in India on compute provided under the IndiaAI mission.

These models represent a true full-stack effort. Beyond datasets, we optimized tokenization, model architecture, execution kernels, scheduling, and inference systems to make deployment efficient across a wide range of hardware, from flagship GPUs to personal devices like laptops. Both models are already in production. Sarvam 30B powers Samvaad, our conversational agent platform. Sarvam 105B powers Indus, our AI assistant built for complex reasoning and agentic workflows.

The Sarvam models are globally competitive for their class. Sarvam 105B performs well on reasoning, programming, and agentic tasks across a wide range of benchmarks. Sarvam 30B is optimized for real-time deployment, with strong performance on real-world conversational use cases. Both models achieve state-of-the-art results on Indian language benchmarks, outperforming models significantly larger in size.

This release marks an important milestone for Sarvam. Building these models required developing end-to-end capability across data, training, inference, and product deployment. With that foundation in place, we are ready to scale to significantly larger and more capable models, including models specialised for coding, agentic, and multimodal conversational tasks.

You can experience Sarvam 105B is available on Indus. Both models are accessible via our API at the API dashboard. Weights can be downloaded from AI Kosh (30B, 105B) and Hugging Face (30B, 105B). If you want to run inference locally with Transformers, vLLM, and SGLang, please refer the Hugging Face models page for sample implementations.

Architecture

Both models share a common architectural principle: high-capacity reasoning with efficient training and deployment. At the core is a Mixture-of-Experts (MoE) Transformer backbone that uses sparse expert routing to scale parameter count without increasing the compute required per token, while keeping inference costs practical. The architecture supports long-context inputs through rotary positional embeddings, RMSNorm-based stabilization, and attention designs optimized for efficient KV-cache usage during inference.

While the two models share the same design philosophy , they differ in scale and attention mechanism. Sarvam 30B uses Grouped Query Attention (GQA) to reduce KV-cache memory while maintaining strong performance. Sarvam 105B extends the architecture with greater depth and Multi-head Latent Attention (MLA), a compressed attention formulation that further reduces memory requirements for long-context inference.

Both models use sparse expert feedforward layers with 128 experts, but differ in expert capacity and routing configuration. This allows the larger model to scale to higher total parameters while keeping active compute bounded.

Training

All stages of the training pipeline were developed and executed in-house. This includes the model architecture, data curation and synthesis pipelines, reasoning supervision frameworks, and reinforcement learning infrastructure. Building everything from scratch gave us direct control over data quality, training dynamics, and capability development across every stage of training, which is a core requirement for a sovereign stack.

Pre-training

Our 30B and 105B models were trained on large datasets, with 16T tokens for the 30B and 12T tokens for the 105B. The pre-training data spans code, general web data, specialized knowledge corpora, mathematics, and multilingual content. After multiple ablations, the final training mixture was balanced to emphasize reasoning, factual grounding, and software capabilities. We invested significantly in synthetic data generation pipelines across all categories. The multilingual corpus allocates a substantial portion of the training budget to the 10 most-spoken Indian languages.

Pre-training was conducted in three phases, covering long-horizon pre-training, mid-training, and a long-context extension phase. We used sigmoid-based routing scores rather than traditional softmax gating, which improves expert load balancing and reduces routing collapse during training. An expert-bias term stabilizes routing dynamics and encourages more uniform expert utilization across training steps. We observed that the 105B model achieved benchmark superiority over the 30B remarkably early in training, suggesting efficient scaling behavior.

Supervised Finetuning

During supervised fine-tuning, the model is trained on a large corpus of high-quality prompts curated for difficulty, quality, and domain diversity. Prompts are sourced from open datasets and labeled using custom models to identify domains and analyze distribution coverage. To address gaps in underrepresented or low-difficulty areas, additional prompts are synthetically generated based on the pre-training domain mixture. Empirical analysis showed that most publicly available datasets are dominated by low-quality, homogeneous, and easy prompts, which limits continued learning. To mitigate this, we invested significant effort in building high-quality prompts across domains. All corresponding completions are produced internally and passed through rigorous quality filtering. The dataset also includes extensive agentic traces generated from both simulated environments and real-world repositories, enabling the model to learn tool interaction, environment reasoning, and multi-step decision making.

For safety fine-tuning, we developed a dataset covering both standard and India-specific risk scenarios. This effort was guided by a unified taxonomy and an internal model specification inspired by public frontier model constitutions. To surface and address challenging failure modes, the dataset was further augmented with adversarial and jailbreak-style prompts mined through automated red-teaming. These prompts were paired with policy-aligned, safe completions for supervised training.

Reinforcement Learning

The reinforcement learning stage uses a large and diverse prompt distribution spanning mathematics, coding, STEM reasoning, web search, and tool usage across both single-turn and multi-turn environments. Rewards are derived from a combination of verifiable signals, such as correctness checks and execution results, and rubric-based evaluations that assess instruction adherence, formatting, response structure, and overall quality. To maintain an effective learning curriculum, prompts are pre-filtered using open-source models and early checkpoints to remove tasks that are either trivially solvable or consistently unsolved. During training, an adaptive sampling mechanism dynamically allocates rollouts based on an information-gain metric derived from the current pass rate of each prompt. Under a fixed generation budget, rollout allocation is formulated as a knapsack-style optimization, concentrating compute on tasks near the model’s capability frontier where learning signal is strongest.

The RL system is implemented with an asynchronous GRPO architecture that decouples generation, reward computation, and policy updates, enabling efficient large-scale training while maintaining high GPU utilization. Trajectory staleness is controlled by limiting the age of sampled trajectories relative to policy updates, balancing throughput with training stability. The system omits KL-divergence regularization against a reference model, avoiding the optimization conflict between reward maximization and policy anchoring. Policy optimization instead uses a custom group-relative objective inspired by CISPO, which improves stability over standard clipped surrogate methods. Reward shaping further encourages structured reasoning, concise responses, and correct tool usage, producing a stable RL pipeline suitable for large-scale MoE training with consistent learning and no evidence of reward collapse.

Benchmarks

Sarvam 105B

Sarvam 105B matches or outperforms most open and closed-source frontier models of its class across knowledge, reasoning, and agentic benchmarks. On Indian language benchmarks, it significantly outperforms all models we evaluated.

General capabilities

Sarvam 105B shows strong, balanced performance across core capabilities including mathematics, coding, knowledge, and instruction following. It achieves 98.6 on Math500, matching the top models in the comparison, and 71.7 on LiveCodeBench v6, outperforming most competitors on real-world coding tasks. On knowledge benchmarks, it scores 90.6 on MMLU and 81.7 on MMLU Pro, remaining competitive with frontier-class systems. With 84.8 on IF Eval, the model demonstrates a well-rounded capability profile across the major workloads expected of modern language models.

Reasoning performance

Sarvam 105B performs strongly on multi-step reasoning benchmarks, reflecting the training emphasis on complex problem solving. On AIME 25, the model achieves 88.3 Pass@1, improving to 96.7 with tool use, indicating effective integration between reasoning and external tools. It scores 78.7 on GPQA Diamond and 85.8 on HMMT, outperforming several comparable models on both. On Beyond AIME (69.1), which requires deeper reasoning chains and harder mathematical decomposition, the model leads or matches the comparison set. Taken together, these results reflect consistent strength in sustained reasoning and difficult problem-solving tasks.

Agentic capabilities

Sarvam 105B is optimized for agentic workloads involving tool use, long-horizon reasoning, and environment interaction. This is reflected in strong results on benchmarks designed to approximate real-world workflows. On BrowseComp, the model achieves 49.5, outperforming several competitors on web-search-driven tasks. On Tau2 (avg.), a benchmark measuring long-horizon agentic reasoning and task completion, it achieves 68.3, the highest score among the compared models. These results indicate that the model can effectively plan, retrieve information, and maintain coherent reasoning across extended multi-step interactions.

Comparison with Larger Models

A useful comparison is within the same scaling regime, since training compute, dataset size, and infrastructure scale increase dramatically with each generation of frontier models. The newest models from other labs are trained with significantly larger clusters and budgets. Across a range of previous-generation models that are substantially larger, Sarvam 105B remains competitive. We have now established the effectiveness of our training and data pipelines, and will scale training to significantly larger model sizes.

Sarvam 30B

Sarvam 30B is designed as an efficient reasoning model for practical deployment, combining strong capability with low active compute. With only 2.4B active parameters, it performs competitively with much larger dense and MoE models across a wide range of benchmarks. The evaluations below highlight its strengths across general capability, multi-step reasoning, and agentic tasks, indicating that the model delivers strong real-world performance while remaining efficient to run.

General capabilities

Sarvam 30B performs strongly across core language modeling tasks, particularly in mathematics, coding, and knowledge benchmarks. It achieves 97.0 on Math500, matching or exceeding several larger models in its class. On coding benchmarks, it scores 92.1 on HumanEval and 92.7 on MBPP, and 70.0 on LiveCodeBench v6, outperforming many similarly sized models on practical coding tasks. On knowledge benchmarks, it scores 85.1 on MMLU and 80.0 on MMLU Pro, remaining competitive with other leading open models.

Reasoning performance

Sarvam 30B performs strongly on multi-step reasoning benchmarks, reflecting its ability to handle complex logical and mathematical problems. On AIME 25, it achieves 88.3 Pass@1, improving to 96.7 with tool use, indicating effective integration between reasoning and external tools. It scores 66.5 on GPQA Diamond and performs well on challenging mathematical benchmarks including HMMT Feb 2025 (73.3) and HMMT Nov 2025 (74.2). On Beyond AIME (58.3), the model remains competitive with larger models. Taken together, these results indicate that Sarvam 30B sustains deep reasoning chains and expert-level problem solving, significantly exceeding typical expectations for models with similar active compute.

Agentic capabilities

Sarvam 30B supports native tool calling and performs consistently on benchmarks designed to evaluate agentic workflows involving planning, retrieval, and multi-step task execution. On BrowseComp, it achieves 35.5, outperforming several comparable models on web-search-driven tasks. On Tau2 (avg.), it achieves 45.7, indicating reliable performance across extended interactions. SWE-Bench Verified remains challenging across models; Sarvam 30B shows competitive performance within its class. Taken together, these results indicate that the model is well suited for real-world agentic deployments requiring efficient tool use and structured task execution, particularly in production environments where inference efficiency is critical.

Indian Language Performance

To evaluate Indian language capabilities, we developed a new benchmark using a pairwise comparison framework with an LLM-as-judge protocol. A key goal of this benchmark is to reflect how language is actually used in India today. This means evaluating each language in two script styles, native script representing formal written usage and romanized Latin script representing colloquial usage commonly seen in messaging and online communication.

The benchmark is organized into four domains: general chat, STEM, mathematics, and coding. It originates from 110 English source prompts, with 50 covering general chat and 20 each for STEM, mathematics, and coding. Each prompt is translated into 22 scheduled Indian languages and provided in both native and romanized script.

Evaluating correctness for complex reasoning prompts directly in low-resource languages can be noisy and inconsistent. To address this, we generated high-quality reference answers in English using Claude Opus 4, which are used only to evaluate the usefulness dimension, covering relevance, completeness, and correctness, for answers generated in Indian languages.

The evaluation uses a pairwise comparison methodology with Gemini 3 as the judge model. The judge evaluates responses across four dimensions: fluency, language/script correctness, usefulness, and verbosity. The evaluation dataset and corresponding prompts are available here.

Sarvam 105B wins on average 90% across all benchmarked dimensions and on average 84% on STEM. math, and coding.

Sarvam 30B wins on average 89% of comparisons across all benchmarked dimensions and 87% on STEM, mathematics, and coding.

Tokenizer Efficiency

The Sarvam tokenizer is optimized for efficient tokenization across all 22 scheduled Indian languages, spanning 12 different scripts, directly reducing the cost and latency of serving in Indian languages. It outperforms other open-source tokenizers in encoding Indic text efficiently, as measured by the fertility score, which is the average number of tokens required to represent a word. It is significantly more efficient for low-resource languages such as Odia, Santali, and Manipuri (Meitei) compared to other tokenizers. The chart below shows the average fertility of various tokenizers across English and all 22 scheduled languages.

Inference Optimization

Sarvam 30B

Sarvam 30B was built with an inference optimization stack designed to maximize throughput across deployment tiers, from flagship data-center GPUs to developer laptops. Rather than relying on standard serving implementations, the inference pipeline was rebuilt using architecture-aware fused kernels, optimized scheduling, and disaggregated serving.

Microsecond-level profiling of the execution stack identified memory stalls, kernel launch overhead, and inefficient scheduling as primary bottlenecks. Addressing these yielded substantial throughput improvements across all hardware classes and sequence lengths. The optimization strategy focuses on three key components.

• Kernel-level rewrites using fused attention and matmul pipelines tailored for each hardware target
• Advanced scheduling and batching strategies that improve GPU utilization under realistic multi-user loads
• Disaggregated serving pipelines that remove bottlenecks between prefill and decode stages

These optimizations yield significantly higher tokens per second per GPU at the same latency targets, enabling higher user concurrency and lower infrastructure costs.

High-End Server Performance (H100)

On H100-class infrastructure, Sarvam 30B achieves substantially higher throughput per GPU across all sequence lengths and request rates compared to the Qwen3 baseline, consistently delivering 3x to 6x higher throughput per GPU at equivalent tokens per second per user operating points.

Performance on cost-efficient deployments (L40S)

Sarvam 30B runs efficiently on mid-tier accelerators such as L40S, enabling production deployments without relying on premium GPUs. Under tighter compute and memory bandwidth constraints, the optimized kernels and scheduling strategies deliver 1.5x to 3x throughput improvements at typical operating points. The improvements are more pronounced at longer input and output sequence lengths (28K / 4K), where most real-world inference requests fall.

Edge Performance (MacBook Pro with MXFP4)

Sarvam 30B is also optimized for local execution on Apple Silicon systems using MXFP4 mixed-precision inference. On MacBook Pro M3, the optimized runtime achieves 20 to 40% higher token throughput across common sequence lengths. These improvements make local experimentation significantly more responsive and enable lightweight edge deployments without requiring dedicated accelerators.

Sarvam 105B

Sarvam 105B is optimized for server-centric hardware, following a similar process to the one described above with special focus on MLA (Multi-head Latent Attention) optimizations. These include custom shaped MLA optimization, vocabulary parallelism, advanced scheduling strategies, and disaggregated serving. The comparisons above illustrate the performance advantage across various input and output sizes on an H100 node.

Tokenizer and Inference Optimization

Combined with the efficient Indic tokenizer, the performance delta increases significantly for the same SLA. For the 30B model, the delta increases by as much as 10x, reaching performance levels previously not achievable for models of this class on Indic generation.

Note: performance numbers are standalone model measurements without disaggregated inference.

Demos

The following demonstrations show the practical capabilities of the Sarvam model family across real-world applications, spanning webpage generation, multilingual conversational agents, complex STEM problem solving, and educational tutoring. The examples reflect the models’ strengths in reasoning, tool usage, multilingual understanding, and end-to-end task execution, and illustrate how Sarvam models can be integrated into production systems to build interactive applications, intelligent assistants, and developer tools.

Webpage creation

The widgets below demonstrate Sarvam 105B’s agentic capabilities through end-to-end project generation using a Claude Code harness, showing the model’s ability to build complete websites from a simple prompt specification.

Pokédex

The goal was to generate a complete, production-ready webpage including all HTML, CSS, and JavaScript required to run the application without frameworks or build tools. The model used the PokéAPI to dynamically load Pokémon data, implementing pagination, search, filtering, and a detailed modal view, all from the prompt shown below.

Product Landing Page

The task was to build a complete website for Sarvam, capturing the spirit of an Indian AI company building for a billion people while matching a world-class visual standard across typography, motion, layout, and interaction design. The full prompt is shown below.

JEE Mains 2026

Sarvam 105B was evaluated on the JEE Main 2026 paper from Shift 2, conducted on 28 January 2026, to demonstrate its STEM reasoning capabilities. The question paper and solutions were sourced from: https://allen.in/jee-main/january-2026-question-paper-with-solutions

The evaluation was carried out in two phases:

1. Text-Only Evaluation: For text-only questions, Sarvam 105B was evaluated directly on questions containing purely textual content.
2. Diagram-Based Evaluation: For questions that included diagrams, Gemini-3-Pro was used to generate structured textual descriptions of the visuals, which were then provided as input to Sarvam 105B for answer generation.

The tables below summarize Sarvam 105B’s performance across Physics, Chemistry, and Mathematics under Pass@1 and Pass@2 evaluation settings.

Under Pass@1, the model shows strong first-attempt accuracy across all subjects. In Mathematics, it achieves a perfect 25/25. In Chemistry, it scores 23/25, with near-perfect performance on both text-only and diagram-derived questions. Physics shows similarly strong performance at 22/25, with most errors occurring in diagram-based reasoning.

Under Pass@2, performance improves to perfect scores across all subjects. Physics improves from 22/25 to 25/25, Chemistry from 23/25 to 25/25, and Mathematics maintains a perfect 25/25. Diagram-based questions in both Physics and Chemistry achieve full marks at Pass@2, indicating that the model reliably resolves visual reasoning tasks when given structured textual representations.

Tutor Mode

Tutor Mode is an internal project where the Indus stack operates with a system prompt optimized for student-teacher conversations. The example below shows Sarvam 105B helping a student solve a JEE problem through interactive dialog rather than providing the answer directly. The model guides the student by asking probing questions, building toward the underlying concepts before arriving at the answer. This also demonstrates the model’s role-playing ability.

Note: the questions below are taken from the same JEE Mains paper solved above.

Codeforces

The coding capabilities of Sarvam 30B and Sarvam 105B were evaluated using real-world competitive programming problems from Codeforces (Div3, link). The evaluation involved generating Python solutions and manually submitting them to the Codeforces platform to verify correctness. Correctness is measured at pass@1 and pass@4 as shown in the table below.

Answers are generated using the following system prompt, with code snippets extracted from markdown fences and think tokens stripped from within tags.

The Codeforces contest used for this evaluation took place in February 2026, while the knowledge cutoff of both models is June 2025, making it unlikely that the models had seen these questions. Strong performance in this setting provides evidence of genuine generalization and real problem-solving capability.

Samvaad: Conversational Agents

Sarvam 30B has been fine-tuned for production deployment of conversational agents on Samvaad, Sarvam’s Conversational AI platform. Compared to models of similar size, it shows clear performance improvements in both conversational quality and latency.

Key strengths include strong proficiency in Indian languages, particularly accurate handling of numerical information within those languages, and reliable execution of tool calls during multilingual interactions. Latency gains come from a combination of fewer active parameters than comparable models, targeted inference optimizations, and reduced tokenizer overhead.

The two examples below show telephonic conversations handled by Sarvam 30B in Hindi and Tamil.

Indus: AI Assistant for India

Sarvam 105B powers Indus, Sarvam’s chat application, operating with a system prompt optimized for conversations. The example demonstrates the model’s ability to understand Indic queries, execute tool calls effectively, and reason accurately. Web search is conducted in English to access current and comprehensive information, while the model interprets the query and delivers a correct response in Telugu.

{
  "query": "pickleball courts Vijayawada Benz Circle Andhra Pradesh",
  "search_type": "general"
}

1. Top Pickleball Courts in Vijayawada near me
2. The Pickleball Republic - Siddhartha Nagar, Vijayawada
3. PickleBall Arena (@pickleballarena_vijayawada)
4. Associations Of Sports in Benz Circle, Vijayawada - Justdial
5. Sports Venues in Benz-circle-vijayawada: Book Top ...

{
  "query": "pickleball equipment cost India beginner paddle shoes racket",
  "search_type": "general"
}

1. Buy Pickleball Equipment Paddles, Balls, Nets Online in ...
2. Buy Pickleball Paddles Online at Best Prices In India
3. Pickleball Equipment
4. Buy Pickleball Paddles Online in India at Best Prices
5. Buy HEAD Pickleball Paddle at Best Price in India

{
  "query": "pickleball beginner rules tips common mistakes how to play",
  "search_type": "general"
}

1. 15 Common Pickleball Errors Ruining Your Game
2. How to Play Pickleball: 9 Rules Every Beginner Should Know
3. 5 common beginner mistakes in pickleball
4. Common Pickleball Mistakes: 5 Errors Beginners Make
5. How to Play Pickleball: The Ultimate Guide on Pickleball Rules

{
  "query": "pickleball court rental price Vijayawada hourly rate",
  "search_type": "general"
}

1. 🏓 Play Pickleball at the Lowest Price Ever in VIJAYAWADA ...
2. Dink It Pickleball - Vijayawada - Guru Nanak Colony ...
3. Pickleball in Vijayawada! Play at The Pickleball Republic
4. 🏓 Play Pickleball at the Lowest Price Ever in VIJAYAWADA ...
5. 5️⃣0️⃣0️⃣ 1 month swimming pool(including training)+ ...

Conclusion

Sarvam 30B and Sarvam 105B represent a significant step in building high-performance, open foundation models in India. By combining efficient Mixture-of-Experts architectures with large-scale, high-quality training data and deep optimization across the entire stack, from tokenizer design to inference efficiency, both models deliver strong reasoning, coding, and agentic capabilities while remaining practical to deploy.

A defining strength of the Sarvam model family is its investment in the Indian AI ecosystem, reflected in strong performance across Indian languages, tokenization optimized for diverse scripts, and safety and evaluation tailored to India-specific contexts. Combined with Apache 2.0 open-source availability, these models serve as foundational infrastructure for sovereign AI development.

This release also marks a milestone in internal capabilities. Through this effort, Sarvam has developed the know-how to build high-quality datasets at scale, train large models efficiently, and achieve strong results at competitive training budgets. With these foundations in place, the next step is to scale further, training significantly larger and more capable models.

Acknowledgements

These models were trained using compute provided through the IndiaAI Mission, under the Ministry of Electronics and Information Technology, Government of India. Nvidia collaborated closely on the project, contributing libraries used across pre-training, alignment, and serving. We’re also grateful to the developers who used earlier Sarvam models and took the time to share feedback. We’re open-sourcing these models as part of our ongoing work to build foundational AI infrastructure in India.