An AI pioneer is launching a blockchain-powered marketplace for open-source models and applications

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week here.

AI pioneer Illia Polosukhin offers a blockchain-flavored open-source enablement platform

Illia Polosukhin, who coinvented the Transformer architecture at Google in 2017, is now launching a new company called Near AI. The company will offer services through a blockchain-powered platform that functions as a secure marketplace for open-source AI models and agents.

Polosukhin is a strong advocate for transparent, open-source AI models, which he believes are the best way to promote responsible adoption in an equitable and safe manner. He’s deeply concerned about individuals’ data privacy and the data sovereignty of companies and countries in a world where just a few tech giants control the largest AI models and the vast amounts of data they collect. His vision is to empower many developers to build powerful models openly, while giving users more control over their data.

The challenge, however, is that developers of open-source (or open-weight) models typically offer their work for free. This means they incur development costs without generating revenue—and must also worry about potential liability if someone later uses their models to cause harm.

According to Polosukhin, some open-source developers are reluctant to host their models on Hugging Face due to the risk of weight leaks. Hosting models on their own server clusters shifts the burden of governance and compliance—especially regarding user data—entirely onto them. On the Near platform, which Polosukhin and his team have been developing since mid-2024, that burden is eliminated. “For developers it’s a very effective way for them to ship their models and agents because they don’t need to deal with all the GDPR compliance themselves for user data,” Polosukhin says. “We build a very simple hosting platform that does all that and they also get distribution.” He likens the Near platform to an app store—but for open-source models and agents.

The platform leverages blockchain technology to create a decentralized network of users—including enterprises, independent developers, and researchers—around the world, Polosukhin explains. Each user must have the appropriate hardware configuration to join the network. Once connected, they can run encrypted versions of models locally and pay a fee per token to the model’s creator. User registration and payment coordination are handled on the blockchain.

Users benefit by gaining access to models from verified developers on a trusted platform—one where neither Near nor the model creators can see the user data running through the models. For enterprises in regulated industries, this could be more appealing than sending sensitive data to a frontier model provider like OpenAI or Google, where privacy and security are less transparent.

Between 2018 and 2022, the Near Foundation raised $650 million to build the Near protocol and platform. Now, Near AI—the newly formed company—will offer specialized services for developers and users on top of that platform. Don’t be surprised if Near AI raises a new funding round in the near future.

With new Google and DeepSeek models, AI parity seems very real

Remember when OpenAI’s Sam Altman said that other frontier model developers shouldn’t bother trying to catch up to OpenAI? What a difference a year makes. For a long time, OpenAI’s models enjoyed a comfortable performance lead over the competition—but that’s changed. Google, which had been slow to accelerate development of generative models that might someday exceed human intelligence, yesterday released a new reasoning model called Gemini 2.5 Pro Experimental that outperforms OpenAI’s best across several well-known benchmark tests.

Gemini 2.5 Pro Experimental now leads in the math and science benchmarks Graduate-Level Google-Proof Q&A (GPQA) and the 2025 American Invitational Mathematics Examination (AIME). It also achieves a state-of-the-art score of 18.8% on Humanity’s Last Exam, a notoriously difficult test set designed to capture the human frontier of knowledge and reasoning.

Google DeepMind leader Demis Hassabis says on X that Gemini 2.5 “reasons through its thoughts before responding, effectively mimicking how humans process thoughts.” “It approaches a problem gradually, refines potential solutions, and chooses the best one,” he says. The model’s standout features include native multimodality—it processes images, video, audio, and code as tokens—and the ability to hold vast amounts of information (up to a million tokens) in memory while reasoning through problems and selecting the best answer.

The plot thickens. The Chinese AI lab DeepSeek has just released a new open-source model called DeepSeek-V3-0324, which improves on the earlier V3 model in reasoning, math, science, and coding. Anyone can download the model from Hugging Face and start developing their own models and apps with it. Developers can’t do the same with OpenAI’s and Google’s top models, which do not reveal their code, training data, or model weights. A year ago, most industry observers believed China was years behind U.S. AI companies in model performance; now, the gap is said to be just a few months.

The takeaway is that state-of-the-art frontier models aren’t as precious or untouchable as they once were. That’s why OpenAI, Google, and others are racing to get consumers hooked on their chatbots, agents, and other applications as quickly as possible. Their main value is no longer in powering enterprise systems at premium prices via APIs—it’s in getting as many consumer users as possible into subscription plans, as AI becomes a bigger part of everyday work and life.

Anthropic and Databricks team up to service enterprises in regulated industries

Databricks and Anthropic are joining forces in a five-year partnership that will integrate Anthropic’s Claude AI models directly into the Databricks Data Intelligence Platform. As a result, Databricks’ more than 10,000 enterprise customers will be able to send their proprietary data through the most advanced Anthropic models within the AWS, Azure, or Google clouds. Many of Databricks’ customers are enterprises in regulated industries such as finance and healthcare, so being able to run advanced AI models on the same platform as their data storage is a major advantage.

Databricks notes that its tools allow enterprises to create AI agents that not only understand their organizational data but can also reason about it using Anthropic’s Claude models. “They are the leading AI model out there and so getting that in front of our customers is really what all this is about,” says Databricks VP of generative AI Naveen Rao. “We want to give our customers the best AI capabilities to unlock the value of their data.”

Databricks customers will have access to Anthropic by default within the Databricks cloud, eliminating the often-protracted process of adding a new vendor. The Anthropic fees will simply appear on their Databricks bill.

Rao says the two companies will first focus on helping existing Databricks customers get the most out of the Anthropic models. They may then shift focus to attracting new enterprise clients with the combined strengths of Databricks’ data security and governance and Anthropic’s state-of-the-art intelligence and safety controls.

“We meet customers all the time who have clarity on their model story but are looking for something in terms of their long-term data story, and I’m sure the opposite or the compliment is true on the Databricks side,” says Anthropic chief product officer Mike Krieger. “You’re often meeting a customer on some portion of their journey, and being able to help them fast-forward through some discovery on the other parts helps you show up well.”