Adept

What Adept is

Adept is an artificial intelligence company founded in 2022^[4]. Two and a half years after starting, it described its mission as building useful general intelligence that lets people and computers work together^[1]. In practice the company worked on two things at once: multimodal foundation models that handle text and images together, and agent software built on top of them.

The original plan

Adept's stated plan was to train progressively larger multimodal foundation models, fine-tune those models into agents, and build products around those agents^[1]. An agent, in this sense, is a system that takes a goal or an instruction and carries out the steps of a task, rather than only answering a single question. The plan therefore combined frontier model training with an enterprise software business — two activities with very different costs and timelines.

Fuyu-8B

On 17 October 2023 Adept open-sourced Fuyu-8B, a multimodal model it described as a decoder-only multimodal transformer with no image encoder^[2]. A decoder-only transformer is the stacked-layer design used by most text-generating models: it produces output one token at a time. Many multimodal designs attach a separate image encoder that turns a picture into features before the language model sees them; Fuyu-8B skips that component and projects image patches directly into the model's first layer^[2].

Because the model drops image-specific position embeddings — the signals that normally record where a piece of the image sits — it accepts arbitrary image resolutions at both training and inference^[2]. That matters for agents, which often have to read screenshots, documents and interfaces whose shapes are not fixed in advance.

Fuyu-8B was shipped as a raw base model under the CC-BY-NC 4.0 licence, which rules out commercial use^[2][3]. Adept said the larger Fuyu-Medium model would not be released^[2].

Funding

Before the Amazon deal, Adept had raised more than 415 million US dollars from backers including NVIDIA, Atlassian, Workday and Greylock, and was valued at about 1 billion dollars after a March 2023 round^[4].

The 2024 change

In June 2024 Adept announced that its co-founders and part of its team were joining Amazon's AGI organization^[1]. Amazon also licensed Adept's agent technology, its family of multimodal models and a few datasets^[1]. Adept said building both foundation models and an enterprise agent product would have required heavy fundraising, so it refocused entirely on solutions that enable agentic AI^[1]. Zach Brock, head of engineering, took over as chief executive, while Tim Weingarten remained head of product^[1]. Chinese state media reported the move in November 2024, noting that the company had been founded in 2022^[4].

Where it sits

Adept's trajectory is a compact example of a problem that faced many model companies: training frontier models and shipping an enterprise product at the same time demands capital on a scale that is hard for a startup to carry for long. The company's open release also left a public artifact — an architecture that avoided a separate image encoder and handled arbitrary resolutions^[2], distributed under a non-commercial licence^[3]. Its later position is narrower: rather than training its own foundation models for products, it builds solutions that enable agentic AI^[1].