Lightspeed Venture Partners is a venture capital firm that engages in consumer, enterprise, technology, and cleantech markets. It focuses on seed, early-stage, later-stage, expansion-stage, start-ups, growth companies, and incubation, and specializes in debt financing for start-ups and growth companies. Lightspeed Venture Partners invests in various sectors that include enterprise, consumer, big data, bitcoin, enterprise technology, cleantech, mobile, internet, financial technology, cloud solutions, e-commerce, storage, media, networking, energy, and software, software-as-a-service, information technology, biotechnology, and social. It mainly invests in countries such as the United States, Europe, Israel, China, and India.

Lightspeed Venture Partners

Peak XV Partners (Formerly Sequoia Capital India & SEA) is a venture capital and growth investment firm investing across India, South East Asia, and beyond. Peak XV has grown to manage over $9bn in capital across 13 funds and invested in over 400 companies, of which ~40 companies have achieved revenues of over $100MM.

Peak XV Partners

Khosla Ventures invests in companies that are bold, early, and impactful. The firm was started in 2004 by Vinod Khosla, co-founder of Sun Microsystems, to provide venture assistance to entrepreneurs. Khosla Ventures invests in a range of areas including AI, climate, sustainability, enterprise, consumer, fintech, digital health, medtech and diagnostics, therapeutics, and frontier technology.

Khosla Ventures

Sarvam AI

Sarvam AI is on a mission to lead transformative AI research that will significantly improve the robustness, performance, and cost-effectiveness of GenAI app development, deployment, and distribution in India.

<div>Based in India, Sarvam AI is engaged in developing language models specifically tailored to support Indian languages. The primary goal of Sarvam AI is to spearhead AI research aimed at bolstering the efficiency and affordability of GenAI applications' development, deployment, and distribution within India. Moreover, the company aims to lead efforts in large-scale data curation for public-good purposes.</div>

Foundation Models

Sarvam AI Launches Indic Language Model ‘Sarvam-1’

Infosys

To develop small language models for Infosys' AI solutions in banking and IT operations, leveraging Sarvam AI's AI infrastructure and NVIDIA's technology stack.

Fine-tuned versions of large language models, which are focused on specific tasks and applications (e.g. code generation, medical event prediction, financial NLP tasks, etc). These are usually developed by training LLMs using custom data and also serve as foundation models that can be developed even further. These models typically use text inputs and can generate text and/or code outputs.

Microsoft

Microsoft partners with Indian startup Sarvam AI to develop voice-based generative AI applications, focusing on Indian languages and context.

Foundation models that can handle a range of natural language processing (NLP) use cases. Typical tasks include generating summaries, translating, answering questions, and classifying text. These models typically use text inputs and generate text outputs.

<div>
 
 Based in India, Sarvam AI is engaged in developing language models specifically tailored to support Indian languages. The primary goal of Sarvam AI is to spearhead AI research aimed at bolstering the efficiency and affordability of GenAI applications' development, deployment, and distribution within India. Moreover, the company aims to lead efforts in large-scale data curation for public-good purposes.
 
 
 Through the OpenHathi series, Sarvam AI is contributing to the AI ecosystem by providing open models and datasets. This initiative, developed in collaboration with academic partners at AI4Bharat, aims to stimulate innovation in Indian language AI by leveraging contributed language resources and benchmarks. Additionally, Sarvam AI plans to introduce enterprise-grade models on its GenAI platform, set for imminent launch.
 
 
 As of December 2023, Sarvam's initial open-source model operates across the top 10 Indian languages and is yet to be launched. This model will be available for trial use by developers, startups, businesses, and government-owned organizations in India.
 
 
 Funding and financials
 
 
 In December 2023, Sarvam 
 <a href="https://sp-edge.com/updates/24761">
 raised USD 41 million
 </a>
 in Series A funding led by Lightspeed Venture Partners to develop fundamental AI systems that are open source, along with creating tools that enable the development of applications, given the limited computing resources accessible to developers in India.
 
 
 Key customers and partnerships
 
 
 In February 2024, 
 <a href="https://sp-edge.com/updates/26477">
 Sarvam AI partnered with Microsoft 
 </a>
 to enable Sarvam AI's Indic voice-oriented large language model (LLM) to be available on Azure, Microsoft's cloud platform.
 
 
 Key competitors: None comparable 
 (No comparable companies in coverage that develops LLMs for Indian languages.)
 
</div>

<div>
 
 AI21 is an Israel-based AI research lab developing large language models (LLMs). The firm also builds GenAI applications based on its LLMs, marketing them to businesses and consumers.&nbsp;
 
 
 AI21 launched its first AI model, &ldquo;Jurassic-1,&rdquo; in August 2021, offering two versions. The larger variant, Jurassic-1 Jumbo, contained 178 billion parameters&mdash;three billion more than OpenAI&rsquo;s 175-billion parameter GPT-3 model. Similar to GPT-3, Jurassic is able to write emails and articles, summarize text, compose poetry and recipes, create website layouts, and generate code.&nbsp;
 
 
 The firm concurrently launched AI21 Studio, a platform allowing users to train custom versions of Jurassic-1 and embed them into products via an API.&nbsp;
 
 
 In May 2022, AI21 confirmed it was working on a new AI system, called the &ldquo;modular reasoning knowledge and language&rdquo; or &ldquo;MRKL&rdquo; framework (pronounced &ldquo;miracle&rdquo;). MRKL augments LLMs, connecting to external sources and symbolic reasoning models to enable the generation of accurate historical information, news on current events, and dynamic information, such as weather, exchange rates, and Wikipedia information. This allows models to accurately answer math queries by calling a calculator and navigate more complex questions by breaking down queries into parts. An initial implementation of MRKL, Jurassic-X, is available on AI21 Studio, which allows users to connect LLMs to their own databases and navigate their data using natural language.
 
 
 In March 2023, the firm launched &ldquo;Jurassic-2,&rdquo; a successor to its initial language model. The new LLM differs from Jurassic-1 and most popular LLMs, in that it incorporates semantic reasoning into its text generation processes alongside statistical inference, enabling the model to more reliably and accurately respond to queries.
 
 
 AI21 says this was achieved by training Jurassic-2 in two phases: A traditional self-supervised phase using an extensive corpus of unstructured and unlabeled data and a second phase, involving a large volume of labeled data to teach the model to follow instructions. Jurassic-2 did not initially implement the MRKL framework, although AI21 confirmed it as an upcoming feature.
 
 
 The company claims that its foundation models are 30% more efficient than those of other providers and it also does not charge additional fees for the use of fine-tuned models.
 
 
 
 
 In March 2024,
 <a href="https://sp-edge.com/updates/27995">
 AI21 launched Jamba,
 </a>
 a model that integrates the Mamba Structured State Space model (SSM) technology and transformer components. The company further introduced 
 <a href="https://sp-edge.com/updates/32879">
 Jamba 1.5 Mini and Jamba 1.5 Large
 </a>
 to offer high performance and efficiency for long-context language processing in August 2024. The models have a hybrid architecture combining Transformer and Mamba approaches, allowing for quality responses with large context windows.
 
 
 Key customers and partnerships
 
 
 As of June 2023, AI21&rsquo;s customers include Latitude, Easyway, Verb, Tweet Hunter, and Harambee.
 
</div>

AI21 Labs

UK-based Stability AI is actively involved in building AI applications and offers a range of foundation models, including:Text-to-image modelStable Diffusion (SDXL 1.0): The open-source text-to-image model incorporates a dual network architecture with a 3.5-billion parameter base model and a 6.6-billion parameter refiner. The model was trained on image-caption pairs from the LAION-5B dataset derived from Common Crawl data. It excels in high-quality image generation, spatial object positioning, and complex concept recognition. The model can be fine-tuned for specific tasks using techniques such as embeddings, hypernetworks, and tools like DreamBooth, a deep-learning generation model.<a href="https://sp-edge.com/updates/30420">Stable Diffusion 3 Medium:</a> a text-to-image open-source AI models with 2 billion parameters utilizing a Diffusion Transformer Architecture and can be fine-tuned using small datasets.<a href="https://sp-edge.com/updates/32260">Stable Fast 3D:</a> Text-to-image model generates 3D images from a single 2D image.In November 2023, the company launched <a href="https://sp-edge.com/updates/23352">two new products</a> Sky Replacer and Stable 3D. Sky Replacer lets users manipulate the aesthetic of the sky in their photos from a choice of nine alternatives, enhancing the overall image appeal and Stable 3D aims to simplify the process of creating textured 3D objects, from images or text prompts, into a quick, automatic activity.Text-to-text modelsStable Beluga 1: The instruction-following language model is based on LLaMA 65 billion and has reasoning capabilities.Stable Beluga 2 (FreeWilly): An expansion to Stable Beluga 1, this model is based on LLaMA 70 billion and reportedly ranks near the top of the Open LLM Leaderboard.StableLM is an open-source language model suite; the company has released alpha versions of three billion- and seven-billion parameter models, with 15 to 65 billion-parameter models to follow. It's designed to generate text and code, and can power various downstream applications. <a href="https://sp-edge.com/updates/25750">Stable LM 2 1.6B, </a>an AI language model designed for multiple languages, was released in January 2024.The Japanese StableLM Alpha, is a seven-billion parameter Japanese language model, which claims to have superior performance compared to other publicly available Japanese language models. It offers text generation capabilities and is released under the Apache License 2.0 for commercial use, with an additional research version available for research purposes.StableCode has three billion parameters and is specifically designed to help programmers with coding tasks by providing autocomplete suggestions and responses to programming instructions.Text-to-audio modelStable Audio, which has 907 million parameters, uses a latent diffusion model architecture for audio and is conditioned on various factors including text metadata, audio file duration, and start time. This conditioning allows control over the content and length of the generated audio.Text-to-image model<a href="https://sp-edge.com/updates/26910">Stable Diffusion 3,</a> which has 800 million range and 8 billion parameters and combines a diffusion transformer architecture and flow matching. Additionally, the model can depict multi-subject prompts, create images, and correct spelling. <a href="https://sp-edge.com/updates/30068">Public preview</a> of this model was announced in June 2024.<a href="https://sp-edge.com/updates/26645">Stable Cascade,</a> a text-to-image model developed with a three-phase approach for image creation and modification.Text-to-video model<a href="https://sp-edge.com/updates/24289">Stable Video Diffusion</a>, an AI model designed to produce videos from images, offered as a research preview. The models, SVD and SVD-XT, can generate short, high-quality clips but come with limitations.<a href="https://sp-edge.com/updates/27695">Stable Video 3D (SV3D), </a>a GenAI tool to create 3D videos to create multi-view 3D models from a single input image and can produce short video content from image or text prompts.<a href="https://sp-edge.com/updates/31945">Stable Video 4D,</a> a model that enables users to upload video and obtain dynamic videos from different views.There has been considerable debate about the ethical and copyright implications of using stable diffusion to create images. The model is particularly permissive when it comes to creating different types of content, which has raised concerns about its potential for abuse. In response to its permissiveness, Stability AI has become involved in copyright infringement lawsuits, with artists and Getty Images raising legal objections to the use of their content in the model's training data.In March 2024, <a href="https://sp-edge.com/updates/27872">Stability AI's founder and CEO, Emad Mostaque, resigned from his role </a>and the company's board to focus on the development of decentralized AI to bring transparency and competitiveness to AI.In April 2024, the company made Stable Diffusion 3 and Stable Diffusion 3 Turbo available on the <a href="https://sp-edge.com/updates/28612">Stability AI developer platform API.</a>In May 2024, Stability AI <a href="https://sp-edge.com/updates/29321">introduced Stable Artisan,</a> a tool to enable media generation for Discord users.Key customers and partnershipsIn December 2022, Stability AI entered a partnership with Amazon, with AWS designated as the preferred cloud provider to facilitate the development and expansion of Stability AI's AI models.In March 2024,<a href="https://sp-edge.com/updates/27403"> Stability AI partnered with Tripo AI,</a> a provider of AI 3D models, to release TripoSR for 3D object reconstruction to cater to the demands of entertainment, gaming, industrial design, and architecture professionals, with responsive outputs for visualizing detailed 3D objects.In April 2024, the company<a href="https://sp-edge.com/updates/28612"> partnered with Fireworks AI,</a> a provider of API platforms, to deliver Stable Diffusion 3 and Stable Diffusion 3 Turbo for enterprises.In September 2024, the company <a href="https://sp-edge.com/updates/33262">launched Stable Image Ultra, Stable Diffusion 3 Large, and Stable Image Core</a>, its text-to-image models on Amazon Bedrock.

Stability AI

Cohere is a Canadian startup founded in 2019 by Aidan Gomez, Ivan Zhang, and Nick Frosst, with backgrounds in natural language processing and prior experience at Google Brain. The firm specializes in the development and deployment of large language models (LLMs) for various applications across enterprise and business use cases. Its foundation model, Cohere Command, is a large natural language text generation model with 52.4 billion parameters, focused on task-oriented dialogue, creating summarization, copywriting, dialogue extraction, and question answering.Cohere also provides Cohere Embed, a model designed to generate vector embeddings that capture the essence of text, improving capabilities in search, content moderation, and intent recognition. These embeddings help empower Cohere's multilingual customizable models, which can be fine-tuned to cater to specific use cases and industries. Additionally, Cohere ensures secure deployment options, including private cloud solutions, for added data security and privacy. In November 2023, Cohere unveiled <a href="https://sp-edge.com/updates/23446">Embed V3</a> with advanced capabilities in matching documents to queries, increasing the efficiency of retrieval augmented generation, and reducing the operational costs of LLM applications. In February 2024, the company introduced <a href="https://sp-edge.com/updates/26583">Aya,</a> an open-source LLM that supports 101 languages.According to Stanford University's HELM benchmarks, Cohere's Command model ranks among the highest-performing models available.In March 2024,<a href="https://sp-edge.com/updates/27502"> Cohere released an advanced language model, “Command-R,</a>” for enterprises. The model offers improved performance on retrieval augmented generation (RAG) and tool usage, with longer context windows of up to 128,000 tokens.In April 2024, <a href="https://sp-edge.com/updates/28198">the company released an advanced language model, Command R+</a>, for business applications. The model offers enhanced retrieval augmented generation (RAG) with citation, supports 10 business languages, and automates workflows. Additionally, <a href="https://sp-edge.com/updates/28463">Cohere also released Rerank 3,</a> a model designed to enhance enterprise search and RAG systems. In the same month, <a href="https://sp-edge.com/updates/28509">the company also introduced a private beta for Cohere Compass,</a> a foundation embedding model designed to allow multi-aspect data indexing and searching, and also <a href="https://sp-edge.com/updates/28789">launched a toolkit available for developers to build AI applications</a> faster, offering production-ready applications that can be deployed across cloud providers, with enhanced security and customization options.In May 2024, Cohere introduced <a href="https://sp-edge.com/updates/29350">Command R Fine-Tuning</a> for enhanced performance for business application workload and<a href="https://sp-edge.com/updates/29761"> released Aya 23, </a>a multilingual language model capable of handling 23 languages,available in 8 billion and 35 billion parameters.<a href="https://sp-edge.com/updates/34141">In September 2024</a>, Cohere released updated versions of its APIs for its AI models Chat, Embed, Rerank, and Classify. These updates, collectively called “API V2,” introduced several upgrades to its existing models, including mandatory model version specification in API calls, consolidation of input parameters for Chat API, and improved tool integration capabilities.&nbsp;Key customers and partnershipsIn November 2021, Cohere partnered with Google Cloud to leverage its infrastructure support and TPUs for product development and deployment. Cohere also partnered with other major cloud providers like AWS and Oracle.In March 2024,<a href="https://sp-edge.com/updates/27307"> Cohere partnered with consultancy firm Accenture</a> to integrate its Command model and retrieval augmented generation (RAG) capabilities to offer GenAI solutions for businesses.In April 2024, <a href="https://sp-edge.com/updates/28932">Cohere partnered with Amazon </a>to incorporate Command R+ and Command R to Amazon's cloud platform, Bedrock.In June 2024, the company<a href="https://sp-edge.com/updates/30492"> partnered with Borderless AI,</a> a provider of AI-powered HR solutions, to create tailored AI agents for HR tasks.The company entered into a <a href="https://sp-edge.com/updates/31052">strategic partnership with i-5O,</a> a provider of AI solutions for manufacturing, as part of Cohere's Startup Program and also partnered with <a href="https://sp-edge.com/updates/31590">Fujitsu</a> in July 2024 to develop a Japanese LLM.In September 2024, Cohere <a href="https://sp-edge.com/updates/33458">partnered with Nomura Research Institute (NRI)</a>, a Japanese consulting firm, to develop the NRI Financial AI Platform.

Cohere

Founded in 2022, Hippocratic AI is a healthcare-focused AI model developed by Health GPT. At its core, Hippocratic AI uses a state-of-the-art (SOTA) large language model (LLM) designed to improve healthcare processes, particularly in non-diagnostic tasks. Hippocratic AI aims to differentiate itself from traditional LLMs by prioritizing pre-training with trusted, evidence-based healthcare content rather than the broader Internet. Claimed to be the first safety-focused LLM for healthcare, the model undergoes a unique Reinforcement Learning with Human Feedback process, known as RLHF-HP, which involves working with healthcare professionals to validate its readiness for use. In addition, the model aims to incorporate a compassionate “bedside manner” in its interactions with patients, recognizing the impact of empathy on emotional wellbeing and health outcomes.According to the company, Hippocratic AI outperformed GPT-4 in majority of tests and certifications. Out of 114 evaluations, it outperformed GPT-4 in 105 cases, with a notable margin of +5% in 74 evaluations and +10% in 43 certifications.Key customers and partnerships&nbsp;In August 2023, the company launched the Hippocratic AI Founding Partner Program, an initiative through which the company collaborates with leading healthcare organizations to become founding partners and actively participate in the development and validation of Hippocratic AI's LLM and technology. The program includes establishing use cases, ensuring clinical and patient safety, and contributing to the advancement of healthcare-specific AI applications. In February 2024, these companies began <a href="https://sp-edge.com/updates/27145">beta</a> testing their LLM for non-diagnostic applications in healthcare.In March 2024, <a href="https://sp-edge.com/updates/27696">Hippocratic AI partnered with NVIDIA</a> to develop AI healthcare agents that have low-latency conversational interactions with patients.

Hippocratic AI

<div>
 
 Founded in 2023, GenHealth.ai is a startup focused on advancing healthcare solutions through the use of GenAI. Its proprietary approach involves training a foundation transformer model, known as DOOG-E, on extensive healthcare data, including 50+ million of coded patient claims and clinical information. DOOG-E is a specialized Generative Pre-Trained Transformer (GPT) AI designed for healthcare applications, with 70 million parameters and one billion tokens. Unlike traditional large language models, DOOG-E is a large medical model (LMM) that uses medical events such as ICD, CPT, RxNorm, and LOINC codes as its &quot;language&quot; to predict future sequences of health events based on historical data.
 
 
 The company claims that its LMM, trained on medical data, is less prone to bias and hallucination compared to its GenAI, large language models (LLM), rendering it more dependable for healthcare applications.
 
 
 Funding and financials&nbsp;
 
 
 In July 2023, the company raised USD 13 million in early funding co-led by Craft Venture and Obvious Ventures. It planned to use the funds to support the development and expansion of its LMM for various healthcare applications, including risk adjustment, care management, and financial benchmarking.
 
 
 
</div>

GenHealth

Foundation models that translate visual data based on features and contextual information identified during training. This enables models to interpret images and video, applying those interpretations to predictive or decision-making tasks. Typical tasks include image recognition and classification into categories. These models typically use image and/or video inputs and generate text outputs.

Foundation models that are trained to understand and work with two or more forms of input data (e.g., text and images) to solve problems and complete tasks using both natural language processing and computer vision. These models can also generate outputs across textual and visual modalities. Typical tasks include visual question answering, image captioning, and generating images based on text descriptions. These models typically use text and/or visual inputs and generate text and/or visual outputs.

Foundation models that use big data to suggest or recommend additional products to consumers. These models are designed to predict user-item interaction probabilities and provide recommendations to users. The models are trained to make these inferences based on historical data on customer interactions and non-interactions. These models typically use user features, item features and user interactions as inputs, and generate numeric values like ratings predictions and matching probabilities as outputs.

Foundation models that are capable of automatic speech recognition, transcription in multiple languages, and translation from these languages to English. These are sometimes referred to as audio models. These models typically use speech/audio as inputs and generate text outputs.

Foundation models that transform user text prompts into video clips.

Foundation models that generate music from text descriptions.

All companies within the space are building large AI models

Companies in this space leverage NLP to train LLMs

The hub exclusively covers companies that have built foundation models

All companies within the space can leverage their foundation models for generative AI applications

The technology enables the creation of personalized content

ANALYST_QUICK_TAKE

PROMINENT_CATEGORY

<div>
 Sarvam AI, an Indian AI startup, has launched Sarvam-1, a significant language model designed predominantly for Indian languages. This model that accommodates around 2 billion parameters covers 10 Indian languages alongside English and can be freely downloaded on Hugging Face's model hub.
 Sarvam-1 deploys a unique tokenizer capable of handling Indic scripts effectively and attaining fertility rates ranging from 1.4 to 2.1 tokens per word. It was trained via Sarvam's 2 trillion - Sarvam-2 - over five days deploying 1,024 GPUs on Yotta's Shakti cluster using NVIDIA's NeMo framework. Interestingly, the company asserts its model's superior performance against more significant models, including Gemma-2 and Llama-3, through benchmarks like MMLU, ARC-Challenge, and IndicGenBench.
</div> Analyst QuickTake: 
 The company raised <a href="https: 
//sp-edge.com/updates/24761"> USD 41 million </a> in December last year to build tailored language models for the Indian market. This February, it partnered with <a href="https: 
//sp-edge.com/updates/26477"> Microsoft </a> to enable Sarvam AI's Indic voice-oriented LLM to be available in its cloud platform Azure.

Sarvam AI

About EDGE

Contact Us

Resources

Area of Focus

Technology Research Made Effortless