Replicate

Replicate makes machine-learning models available through hosted APIs and lets developers deploy their own models on GPU infrastructure. Founded by Ben Firshman and Andreas Jansson in 2019, it developed Cog, an open-source model-packaging tool, and grew with the communities building on generative image and language models. Replicate became part of Cloudflare on December 1, 2025.[3][4][6]

Cog packages machine-learning models in standard Docker containers that can run on Replicate or a developer’s own infrastructure. Jansson brought experience building Docker-based model tools at Spotify; Firshman had created Docker Compose while at Docker. Their approach addressed two related problems: making research code reproducible and turning it into a production service.[4][3]

Cog adds a model interface above Docker’s general-purpose container machinery. In his February 2024 interview, Firshman described a Docker image carrying an OpenAPI specification that defines the model’s inputs and outputs, with access to a CUDA device when needed. Its cog.yaml configuration lets researchers specify Python versions and dependencies without writing a Dockerfile. The aim was to package recurring deployment work—CUDA compatibility, GPU access, inference serving and schemas—behind a higher-level tool.[1]

By February 2023, Replicate offered a library of open-source models callable with a few lines of code, Docker images with a standard API, and deployment of developers’ own models with automatically generated API servers. Firshman later explained why packaging alone had been insufficient: getting a container running on a GPU remained difficult for software engineers, so a hosted API became the practical route to using the models.[3][1]

That combination of containers and hosted APIs followed several earlier experiments. Before Replicate’s 2019 start, Firshman and Jansson had collaborated on tools for reading scientific papers; their subsequent work explored how to share and run the software behind those papers.[1][6]

Firshman recalled that they created arXiv Vanity in 2017 while on vacation in Greece, after Jansson struggled to read a machine-learning paper on his phone. They built LaTeX-to-HTML converters to make papers readable as web pages. They later began integrating the work into arXiv, but a collaborator Firshman identified only as Dan picked up and completed that integration after they became occupied with Replicate.[1]

Jansson’s difficulties applying published models at Spotify supplied a more technical problem: papers, incomplete code and separately stored weights did not reliably produce runnable software. The pair proposed containers for models and a central benchmarking service that would let researchers compare existing work on the same evaluations. Firshman recalled struggling to find a business during their Y Combinator batch, then building COVID-related tools after March 2020. They also tried open-source experiment tracking: that project initially used the Replicate name and was later renamed Keepsake.[1][3]

They returned to model sharing with Cog and Replicate. The initial service let researchers publish a webpage where others could try a model and download its Docker container. It had no public API, and they dropped the proposed central benchmarks as too complicated. According to Firshman, researchers often published a model and then stopped using the service until their next paper, limiting repeat use.[1]

Early-2021 image-generation communities provided a different pattern of use: people continually copied, modified and shared notebooks. Advadnoun’s The Big Sleep and RiversHaveWings’ VQGAN+CLIP generated images by guiding a GAN with CLIP; these were community projects, rather than Replicate inventions. Firshman recalled that Pixray became one of the first models published primarily on Replicate around mid-2021, helping establish its early community.[3][1]

The API business began when a user reverse-engineered the model-demo web form’s internal API to generate images from a script. Rather than block the traffic, Replicate documented the interface in Notion and offered paid access. In Firshman’s account, the first customer was PixRay’s creator, generating NFT art and paying “on the order of a thousand bucks a month.” Replicate subsequently introduced an official API and billing.[1]

Stable Diffusion’s 2022 release expanded both the supply of models and demand from product builders. Firshman recalled community developers publishing animation models, repeatable game-texture generators and personalized fine-tunes on Replicate. The founders later credited their existing serving infrastructure with enabling them to handle that demand. By December 2023, Replicate reported that language models accounted for most of its growth and that Llama 2’s release week had been its largest growth week.[1][6][2]

On December 5, 2023, Replicate reported raising a $40 million Series B, led by a16z with NVentures (NVIDIA), Heavybit, Sequoia and Y Combinator participating. It reported 2 million signups over the preceding year and a half and 30,000 paying customers. These were historical signup and customer counts; Firshman explicitly clarified in the February 2024 interview that the 2 million users were not all developers. Organizations and deployments had also been added as features for business use at scale.[2][1]

Replicate’s December 2023 customer examples included Unsplash using BLIP to label its image catalog, BuzzFeed turning pets into plushies, and Character AI and Labelbox deploying models. Firshman explained that Unsplash’s generated descriptions made images searchable. He distinguished customers running off-the-shelf models from those fine-tuning models or using Replicate as infrastructure for custom Python implementations.[2][1]

In February 2024, Firshman said Replicate’s production workloads primarily ran on public clouds, including GCP and CoreWeave, rather than GPUs it owned. Aggregating customer demand allowed it to make larger commitments to cloud providers while selling smaller units of inference. He also described working with Meta and Stability to maintain and optimize popular models such as Llama 2 and Stable Diffusion, while acknowledging that the long tail of community-published models remained less optimized.[1]

Cloudflare announced an agreement to acquire Replicate on November 17, 2025. That day, Rita Kozlov and Firshman reported a catalog exceeding 50,000 open-source models and fine-tuned models, alongside marketplace access to proprietary models including GPT-5 and Claude Sonnet through a unified API. The catalog figure included fine-tunes, rather than measuring unique model architectures. On December 1, Jansson and Firshman announced that Replicate was officially part of Cloudflare.[5][7][6]

The founders explained the decision through customers’ increasingly complex application stacks: model inference was becoming one component alongside microservices, storage, caching, databases and telemetry. They saw Cloudflare’s Workers, R2 and Durable Objects as complements to Replicate’s model-running tools. Bringing the catalog to Workers AI, adding fine-tuning and custom-model capabilities, and connecting inference more deeply with the application platform were acquisition-era plans.[6][7]