Replicate Raises $40 Million For Its Library Of Open Source AI Models

Amid the AI land grab, open source models whose code is available to anyone to modify and use are catching up to their proprietary counterparts. They’re getting bigger (Meta’s Llama 2 is trained on 70 billion parameters) and even getting better than ChatGPT at performing specific tasks. Now, as interest piques in open source AI tools, it’s a prime time to be a startup that hosts and runs open source software, said Ben Firshman, CEO of Replicate, a platform used by 2 million software developers to access and tinker with more than 25,000 open source AI models.

When popular open source models like text-to-image model Stable Diffusion 2.0 and Meta’s large language model Llama 2, were added to its library this year, Replicate saw its biggest spikes in growth. The platform also saw an uptick in traffic after OpenAI’s leadership went through a sudden but temporary shakeup in late November.

“There’s been a larger interest in people switching to open source models because they don’t want to be locked into a proprietary platform that might disappear at some point,” Firshman said.

The startup announced Tuesday that it has raised $40 million in Series B funding in a round led by Andreessen Horowitz with participation from Nvidia’s VC arm NVentures, Heavybit, Sequoia and Y Combinator. The round values the San Francisco-based startup at a $350 million valuation, according to sources familiar with the matter, and brings Replicate’s total funding to about $58 million.

Firshman and cofounder and CTO Andreas Jansson discovered the need for a platform that would make it easier for software developers to use the latest AI models before today’s biggest open source models, like Meta’s Llama 2 and Stable Diffusion, entered the mainstream.

While working as a machine learning engineer at Spotify, Jansson realized that most advancements in AI are locked up inside academic research, hidden behind lengthy descriptions and complicated diagrams, rendering them useless to solve real world problems. In 2019, he teamed up with former colleague Firshman, who had created a system for developers to package up and ship their work while leading product at software unicorn Docker, to launch Replicate. Their goal was to do the same for researchers by making their open source AI and machine learning software available to others.

“There’s been a larger interest in people switching to open source models because they don’t want to be locked into a proprietary platform that might disappear at some point.”

Ben Firshman, CEO and cofounder of Replicate

Replicate isn’t the only startup providing compute resources to run open source models. Competition comes from both highly-valued startups like Together AI, which recently raised a $102.5 million Series A round, $4.5 billion Hugging Face and OctoML, valued at $850 million, as well as tech giants like Nvidia, Google, Amazon and Microsoft, all of which offer similar products to run and customize machine learning models on the cloud.

One argument against open source models is that there’s a range of safety risks: they can be used for malicious purposes like engineering phishing and biological attacks, but open source advocates argue that having model code be transparent means that it will face more scrutiny, which will ultimately make the models safer. Replicate has partly addressed this issue through filters that detect and restrict models from generating harmful content. But as the checkers are prone to incorrectly flagging safe content as unsafe, they can be disabled.

Yet, an increase in demand for open source models is evident on Replicate’s platform, which is home to models that can generate and edit music, videos, text and images. A “face restoration” AI model that can convert blurry old photos into crisp images has been used about 60 million times. Another AI model that can swap one face for another face within two seconds has run almost 30 million times. That’s partly because large open source AI models can be fine-tuned for specific use cases by training on custom data, making them cheaper and faster to use, Firshman said. “You can fine-tune some of these models for a dollar and 10 minutes,” he said.

But this fine-tuning is also a more complicated process than models that you can use off the shelf. “Open source technology is harder to use than closed source products, almost by definition,” said Matt Bornstein, a partner at Andreessen Horowitz who led the round.

Replicate charges developers for the period of time a model is running, ranging anywhere from 36 cents to $20 per hour. The startup has partnered with NVIDIA to provide GPUs of different sizes and capabilities and works with multiple cloud providers like Coreweave and Google Cloud. “Compared to a lot of other AI companies, we have a very clear business model in that we sell infrastructure in exchange for money,” he said, adding that the startup is not yet profitable.

The new funding will be used to attract more software developers to the platform and provide enterprise customers with additional services like security, compliance and monitoring how a model performs. Among its 30,000 paying customers are companies like Buzzfeed, Getty-owned Unsplash and startups like Character AI and Labelbox use Replicate to run open source models. A software engineer himself who previously founded three tech startups, Firshman admits he’s still not a machine learning expert like his cofounder Jansson. But through Replicate, which simplifies using open source AI models, the technology is more accessible to him and others like him.

“Machine learning models can do a lot of the just annoying bug squashing work that software developers actually spend most of their time doing,” he said. “I can do the fun creative stuff.”