The collaborative machine learning startup FedML Inc. said today it has closed on an $11.5 million seed funding round.

The round was made up of two separate tranches. The first raised $4.3 million and was disclosed in March, while the second pulled in $7.2 million and closed earlier this month. Camford Capital led the round, with other participants including Road Capital, Finality Capital Partners, PrimeSet, AimTop Ventures, Sparkle Ventures, Robot Ventures, Wisemont Capital, LDV Partners, Modular Capital and the University of Southern California.

FedML has created a collaborative machine learning operations platform that empowers companies and developers to work together on machine learning tasks by sharing data, models and compute resources. According to the company, its goal is to create an ecosystem for the collaborative development of customized AI models.

The startup explains that many businesses are interested in training and fine-tuning AI models on their own proprietary datasets, so they can better perform tasks around customer service, product design and business automation. But it’s difficult for them to use this data, which is often regulated or siloed, in a safe way with existing, cloud-based AI training systems.

FedML’s solution is a federated machine learning platform that supports the collaborative training of AI models on private and siloed data at the edge. It means that data does not have to be moved to another location. This “learning without sharing” approach promises to be a game changer for some businesses, enabling a healthcare organization to build AI models capable of detecting rare genetic diseases by training it on highly sensitive data stored at various different hospitals, for example.

The other advantage of FedML’s collaborative approach is it reduces the huge resource demands of AI training. It’s known that OpenAI LP, the company that built ChatGPT, spent millions of dollars training it.

Of course, not every company has the same kinds of financial resources. FedML overcomes the need to acquire expensive graphics processing units by allowing users to train and serve their models in any location, using any kind of hardware.

FedML co-founder and Chief Executive Salman Avestimehr said enterprises ideally want to build their own custom AI models, rather than deploy generic large language models created by the likes of OpenAI and Google LLC.

“Large-scale AI is unlocking new possibilities and driving innovation across industries, from language and vision to robotics and reasoning,” he said. “At the same time, businesses have serious and legitimate concerns about data privacy, intellectual property and development costs. All of these point to the need for custom AI models as the best path forward.”

FedML has made a lot of progress since launching in March 2022, building up a thriving open-source community that numbers more than 3,000 users. It has performed more than 8,500 AI training jobs across 10,000 edge devices, and its open-source federated machine learning library has become the most popular in the industry, surpassing Google’s TensorFlow Federated. The company claims to have secured more than 10 enterprise contracts across industries such as healthcare, retail, financial services, smart homes and cities and mobility.

Its most recent development was the introduction of FedLLM, which is a custom training pipeline for building domain-specific LLMs on proprietary data. It’s said to be compatible with popular LLM libraries such as HuggingFace and DeepSpeed, and developers can get started quickly by adding just a few lines of source code to their applications. FedLLM then takes care of all of the complex steps involved in training, serving and monitoring customer LLM models.

“FedML has a compelling vision and unique technology to enable open, collaborative AI at scale,” said Camford Capital Partner Ali Farahanchi. “In a world where every company needs to harness AI, we believe FedML will power both company and community innovation that democratizes AI adoption.”

Image: Wirestock/Freepik