SUTRA’s dual-transformer approach extends the power of both MoE and Dense AI language model architectures, delivering cost-efficient multilingual capabilities for over 50+ languages. It powers scalable AI applications for conversation, search, and advanced reasoning, ensuring high-performance across diverse languages, domains and applications.
Multilingual
Ultrafast response-generation and instruction-following in 50+ languages. Lightweight Dual2 architecture and purpose-built tokenizer enables engaging and personalized conversational experiences with unparalleled inference efficiency.
Online
Connected and hallucination-free models that provide accurate and up-to-date responses in a conversational tone. Secure, on-premise deployment with seamless integration to structured or unstructured enterprise databases and online results.
Thinking (NEW)
Models that deliver deeper, structured thinking across topics and domains. Building on our multilingual AI advances, reasoning models bring complex decision-making, support multiple languages, and use resources efficiently.
50+ Languages With Accuracy & Efficiency /
SUTRA™ models surpass leading models on the MMLU benchmark in comprehending and generating responses across numerous languages.
SUTRA-R0-Preview is still under active development and is showing significant promise, closing the gap in multilingual MMLU performance to larger models like DeepSeek-R1-600B and OpenAI-o1. The benchmark numbers are reported from the February 3, 2025, checkpoint. Our evaluation methodology follows a rigorous multilingual assessment framework inspired by Lai et al. [2023] and Üstün et al. [2024], with key distinctions. Unlike Okapi’s 25-shot evaluation, we adopt a more challenging 5-shot approach, aligning with Hendrycks et al. [2021]. To ensure broad linguistic coverage, we assess performance across three major language groups: [English] [Korean, Japanese, Arabic] and [Indian languages].
Cost-efficient Tokenization
for non-English Languages /
SUTRA's purpose-built tokenizer is more efficient across languages, enabling compute cost-savings and ultrafast inference for multilingual applications.
Evaluating Tokenizer Performance of Large Language Models Across Official Indian Languages (Published November 24)
An independent study (https://arxiv.org/abs/2411.12240) shows SUTRA outperforms GPT-4o, Llama 3.1, and other Indian language models across 14+ Indian languages—delivering superior performance with better token efficiency and cost-effectiveness.
*Currently supported languages are English (en), Hindi (hi), Gujarati (gu), Korean (ko), Japanese (ja), Bengali (bn), Marathi (mr), Telugu (te), Tamil (ta), Kannada (kn), Malayalam (ml), Punjabi (pa), Odia (or), Urdu (ur), Sanskrit (sa), Assamese (as), Thai (th), Arabic (ar), Persian (fa), Vietnamese (vi), Indonesian (id), Turkish (tr), Polish (pl), Russian (ru), Ukrainian (uk), Dutch (nl), French (fr), Italian (it), Spanish (es), German (de), Greek (el), Hebrew (he), Portuguese (pt). More languages are coming soon.
Up-to-date and
Hallucination-free /
SUTRA-Online are internet connected and hallucination-free models that understand queries, browse the web, and summarize information to provide current answers.
SUTRA-Online can answer queries like “Who won the game last night” or “What’s the current stock price” accurately, whereas offline models suffer from knowledge cut-off dates.
Models /
SUTRA-V2 (NEW)
SUTRA-V1 is our multilingual AI model designed for instruction execution and conversational intelligence across 50+ languages. Trained on a diverse blend of proprietary and open-access datasets, it excels in handling complex tasks with high accuracy. With deep proficiency across Latin, Indic, and Far Eastern languages, SUTRA-V1 delivers natural, context-aware responses, making it a powerful solution for global enterprises, multilingual assistants, and cross-language AI applications.
SUTRA-R0
SUTRA-R0 is the first in our series of advanced reasoning models, designed for complex problem-solving and deep contextual understanding. Built to analyze, infer, and generate logical responses, SUTRA-R0 goes beyond pattern recognition—applying structured reasoning to tackle nuanced queries, multi-step problem-solving, and enterprise decision-making. Its architecture enables high-accuracy responses across domains, making it a powerful tool for knowledge-intensive workflows and next-generation AI applications.
SUTRA-Q0
SUTRA-Q0 is a time-series quantitative AI model built on a dense architecture, designed for forecasting and analytical applications beyond text and chat. Optimized for dynamic data patterns, SUTRA-Q0 enables enterprises to predict trends, detect anomalies, and make proactive decisions across finance, logistics, healthcare, and other data-driven domains. Its high-precision forecasting capabilities make it a powerful tool for businesses seeking deeper insights and future-ready strategies.
SUTRA™ for Developers /
SUTRA models are the technology backbone of TWO’s products and services and are available as Model as a Service (MaaS) to other apps and services via usage-based pricing and simple-to-integrate APIs.
🟢