NEAR AI Launches Research Hub to Build the Next Frontier AI Model

Open Source is broken. If the current scaling curve continues, the next generation of frontier AI models will have more than one trillion parameters. Training such a model will cost hundreds of millions of dollars—several times more than the biggest models today—and the generation after it will cost even more.

Today only large centralized AI labs can train such models. Few like Meta actually release their models in the open. We want to ensure that the best AI models are always available to everyone, permissionlessly. Everyone should have access to frontier AI models with absolute certainty that their data cannot be seen by anyone. Today we're launching the NEAR AI Research Hub, a community-built, decentralized frontier AI lab that can build bigger and better models than even the biggest, best-funded centralized companies can.

Our first goal at NEAR.AI is to create a model with 1.4T parameters (3.4x bigger than Meta's latest Llama model). This requires solving two problems: how to create the best 1.4T parameters model, and how to make it economically viable. It must be monetizable once it's trained; it needs a well-orchestrated decentralized research approach that can compete with the big companies' labs; and it needs to be trainable over a physically distributed and/or slow network.

To get to the best 1.4T parameter model, we are starting a continuous competition on a 0.5B parameter model starting every month. We provide resources to all the researchers who want to contribute ideas towards the goal, including the ways to improve the quality and speed of the training loop, data preparation, and filtering. The results are measured on a publicly available benchmark.

We then continuously choose the subset of participants who performed best in this first level, and have them work on creating the best 2B model, and then the best 8B, 30B, 70B, 350B, and finally 1.4T. At each level, the number of participants gets smaller, and the model size becomes larger, until eventually the best researchers work on the final 1.4T parameter model, equipped with all the ideas and experiments from the previous steps. New researchers can keep joining at the entry level of each 0.5B competition cycle.

The smaller the number of participants, the more it will move from a competitive model to a collaborative one. The number of resources available will also increase over time. Let's expect that if we start with 1000 participants in the first step, then somewhere around the 20 best researchers will work on the finishing touches for the 1.4T.

The first competition to train the 0.5B parameter model is open right now. Learn more, apply to participate, and subscribe for updates about future competitions at app.near.ai.

Trusted Execution Environments

How will we pull off building a 1.4T parameter, monetizable model in a decentralized way? Enter the world of Trusted Execution Environments, or TEEs. Modern processors and GPUs (starting with H100s) allow one actor, Alice, to run code on a machine of another actor, Bob, without trusting Bob, while having the following two guarantees: (a) Bob is in fact running the code Alice expected him to run, and (b) Bob cannot spy on the execution that Alice wants him to carry out. In other words, Alice can safely send any private data into such a computation, being certain Bob will not be able to see it.

This enables many great scenarios, including private inference: a network of TEE-enabled machines can run open source models, and anyone can run inference on their private data on such a network while being 100% certain their data is not seen by others.

TEEs enable training of a frontier large-language model in such a way that people can freely copy it to each other, but cannot use the model for free. The model can have fully open source implementation, with all the raw data and data processing pipelines open. It can then be trained on a cluster of TEE-enabled machines.

In this setup, the model is available to everyone, it can be run locally as long as the user has a TEE-enabled machine, everyone knows and can verify the exact way the model was trained; but each usage of the model must be paid for. We can use payment channels to enable the open source community to continuously pool resources to train larger and larger frontier LLMs, available to everyone in a permissionless fashion, and to be certain that the investment will pay off, fueling a sustainable way to then train each consecutive model.

Today NEAR is actively developing SDKs for TEE-enabled inference and training, and works with multiple ecosystem projects to ensure that a large number of machines with H100s are available on NEAR-enabled GPU marketplaces.

Decentralized Infrastructure

To enable progress on open source frontier AI, we need a single decentralized place where people share, prepare, and monetize data and models, run and share experiments, provide and find compute resources, and generally conduct research. Massive datasets need to be stored, and APIs, real-time data providers, vector store databases, and other services in the decentralized networks all need verifiable liveness.

NEAR is building all the infrastructure components, such as decentralized storage, decentralized compute clusters, and the networking protocol with response and delivery guarantees directly into NEAR Protocol, with help from our ecosystem projects.

Join Us

With TEE-enabled inference and training, an ecosystem contributing the necessary decentralized infrastructure components, and a training competition to build the best frontier model in an open and distributed way, we believe NEAR AI Research Hub is poised to achieve User-Owned AI that is better than the for-profit, centralized AI dominating the market today. Join us at app.near.ai or apply to participate in the competition.