Reka AI - Multimodal agentic reasoner you can deploy anywhere ❇️

CV Deep Dive

Today, we’re talking with Dani Yogatama, Co-founder and CEO of Reka AI.

Reka is an AI research and product company focused on building state-of-the-art multimodal models and AI-powered applications. Founded by Dani and his team, Reka trains its own foundation models and develops products that enable enterprises to integrate AI seamlessly into their workflows. The company is best known for its model releases—Reka Edge, Flash, and Core—as well as its enterprise AI platforms, Nexus and Guardian, which are currently in private beta.

Today, Reka is working with enterprises across finance, media, gaming, ecommerce, and government, helping them automate processes, enhance decision-making, and leverage AI for complex tasks. Their models are designed for high performance, privacy, and cost efficiency, making them a compelling alternative in the rapidly evolving foundation model space.

In this conversation, Dani shares how Reka got started, the breakthroughs they’ve made in multimodal AI and agentic reasoning, and how they’re pushing the boundaries of AI efficiency and usability.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Dani 💬

Dani - welcome to Cerebral Valley! First off, introduce yourself and give us a bit of background on yourself and Reka. What led you to start Reka? 

Hi everyone! I’m Dani Yogatama, CEO and co-founder of Reka. I grew up in Indonesia and received my PhD from Carnegie Mellon University. Prior to Reka, I worked for several years at DeepMind in London. My research focuses on multimodal models, memory systems, and reinforcement learning.

Reka is an AI research and product company. We are a full-stack company that trains our own models and develops applications. Most people know us from our model releases last year—Reka Edge, Flash and Core. On the research side, our focus is on multimodal agentic reasoning. On the product side, we develop Nexus—a platform for organizations to create and manage AI workers—and Guardian, a real-time video monitoring system. Both are currently available in private beta. I cannot wait to share them with the general public soon.

We started Reka because we believe in the transformative power of AI. My primary goal is to ensure that this technology benefits as many people as possible. I believe Reka is the best vehicle for us to make the highest impact. At a personal level, I also wanted to try something different after being in a big tech company for a long time. I am very much enjoying the journey so far.

What sets Reka apart from other players in the foundation model space? 

I do not think that we need tens of billions of dollars to push the boundaries of AI capabilities. We are reasonably well capitalized to pursue our goals, but we do not have as much resource as larger companies that have hundreds of thousands of GPUs to train models. As a result, our focus from the beginning has been on efficient training and serving infrastructure and a small number of focused-research bets.

Today, we have a training infrastructure that allows us to train large-scale multimodal models efficiently. All of our models are trained from scratch for less than $10 million and they perform competitively on benchmark numbers. Most of our algorithmic breakthroughs are in post-training, leveraging our expertise in reinforcement learning and imitation learning.

Talk to us about your users today. Who is finding the most value in Reka? 

Our customers range from large enterprises to small and medium businesses in domains such as finance, media and entertainment, gaming, ecommerce, and government organizations. Some of them benefit from base models that we develop, while others use our end-to-end platform to automate functions within their organizations. They work with us because they care about model performance, privacy, and cost-efficiency.

Walk us through Reka Core, Flash and Edge. Which existing use-case for Reka has worked best?

Our models excel at multimodal understanding—particularly extraction of any multimodal data into structured outputs—and agentic reasoning. They have the capabilities to plan, use tools, verify their execution steps, and learn from user feedback.

All of our public models—Reka Spark, Reka Edge, Reka Flash, and Reka Core have the same set of capabilities. They are trained the same way and all of them have full multimodal inputs. They are only different in model sizes.

In practice, they are used for information extraction from PDF documents, real-time video analysis, generating metadata to improve search from images and videos, and creating an interactive gaming experience, among others.

Given the excitement around new trends in AI such as Agents and Multimodal AI, how does this factor into your product vision for Reka? 

Multimodal has been a core thesis since the beginning of Reka. We are one of the first companies (alongside Google) that offer full multimodal capabilities including text, images, video, and audio. Today, we are proud to serve one of the best multimodal models on the market. I believe our visual understanding capabilities are state-of-the-art.

In terms of agents, I think it is a natural evolution of the technology. The first wave of AI is centered around chatbots. As the models become better, we expect them to do more things for us beyond answering questions. I actually think the term ``agents’’ is heavily overused right now. A lot of people call their AI systems agentic even though they only use orchestration that combines a few models to produce an output. Those systems, while useful, are not agentic systems. For me, an agent has to interact with an environment, take actions, and accomplish tasks.

Last year, we were focusing on producing multimodal agentic base models that serve as a foundation for our product offerings. Over the past several months, in addition to continuing to improve the base models, we have expanded our product team to provide a more seamless end-to-end experience. Our goal is to not only provide base models, but also full applications with delightful user experience to interact with powerful AI. We think there are a lot of opportunities to innovate in this space.

How do you see Reka’s product progressing over the next 6-12 months? 

I think we are entering an exciting time where AI models are starting to be deployed in production to solve many real-world problems. In addition, the cost of developing these models continues to go down. We have shown how and will continue to develop cost-efficient state-of-the-art models to support our products. As a full stack company that has the expertise to train models and develop applications, we are well positioned to deliver the highest values for our customers.

In the short term, we have a few exciting launches. I am excited about introducing Nexus to the public. It is a platform for organizations to create and manage AI workers to automate repetitive tasks and streamline operations. Nexus offers seamless integration with internal data, flexible and secure deployment, and a no-code user interface that supports easy creation of these workers. In addition, we will launch Guardian, a video monitoring product built on top of our models that allows users to detect complex events in real-time. Both of these products are built on top of our proprietary models. 

On the research side, we continue to invest on multimodal agentic reasoning. We focus our efforts on reinforcement learning, self play, imitation learning, and multi-agent learning. We have a few technical breakthroughs that underpin Nexus and Guardian which we intend to share more broadly when we can.

What has been the toughest technical challenge around building Reka into the platform it is today?

Building large-scale training infrastructure. People tend to underestimate the optimization that is needed to do it at scale, whether that is for multimodal data processing, pre-training, or post training. A good and robust infrastructure is critical to iterate on ideas to develop our products quickly. We invested heavily in this at the beginning and are now in a good position to continue to benefit from it.

Lastly, how would you describe the culture at Reka? Are you hiring, and what do you look for in prospective team members?

Yes, we are hiring and have many open roles. I believe that it takes a collection of world-class individuals working together towards a common goal to make meaningful contributions in AI in this era. Beyond the necessary technical skills to contribute to our mission, you need to have the right characters to succeed at a startup. It is a high-pressure environment with lots of ups and downs. Resilience and perseverance are as important as engineering skills.

Anything else you’d like people to understand about the work you’re doing at Reka?

We envision a future where organizations, from small startups to large enterprises, operate efficiently by employing and collaborating with AI workers, allowing human employees to focus on critical decisions, oversights, and social components of running a business.This requires capability improvements in multimodal understanding, planning, and reasoning. We continue to develop our models and product offerings towards this goal. 

If you would like to automate certain functions within your organization with our platforms, contact our sales team. If you are excited about research and building innovative AI products, apply to join our team.

Conclusion

Stay up to date on the latest with Reka, learn more about them here.

Read our past few Deep Dives below:

• Build specialized RAG agents with Contextual AI 🌐

• Remyx - Your AI Production Assistant 💡

• Qdrant's GPU-accelerated vector indexing is here 🔋🌐

• Athena is your AI-powered remote hire 🎛

• Doowii - your AI-first education platform 📖

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs