The Abu Dhabi Lab That Gave Away Its GPT-3 Rival for Free: Inside the Falcon LLM Story

The Lab in Abu Dhabi That Gave Away Its GPT-3 Rival for Free

In March 2023, while the world was still getting used to ChatGPT, a government-backed research institute in Abu Dhabi quietly announced something unusual: a large language model that it claimed matched or beat GPT-3 on several benchmarks, built with a fraction of the compute — and unlike the AI labs in San Francisco, it planned to give the whole thing away.

That model was Falcon, and the institute behind it, the Technology Innovation Institute (TII), has spent the three years since turning a one-off research flex into one of the most-downloaded open-source AI model families on the planet. It's a story that rarely gets told outside AI research circles, but it's arguably the UAE's biggest technical contribution to the global AI race so far.

Who's Actually Behind Falcon

TII was established in 2020 as the applied-research arm of Abu Dhabi's Advanced Technology Research Council (ATRC), a government body tasked with building up the emirate's capability in advanced technology — everything from AI and quantum research to secure systems. Falcon came out of TII's AI and Digital Science Research Center, and its goal from day one wasn't to build a flashy chatbot app. It was to prove that a well-run research lab outside the usual US tech giants could train a frontier-class model and then open it up for anyone to use, modify, or build a business on.

According to TII's own announcement, the first Falcon model — a 40-billion-parameter system — was trained on one trillion tokens of data and, by the institute's internal testing, outperformed Meta's then-leading open model and GPT-3, while requiring only around 75% of the training compute GPT-3 had used. Two months later, in May 2023, TII released Falcon 40B and a smaller Falcon 7B under the Apache 2.0 license — a genuinely permissive open-source license, not the restricted "open-weight" terms some rivals used — according to the UAE's state news agency WAM. Falcon 40B promptly climbed to the top of Hugging Face's Open LLM Leaderboard, the closest thing the open-source AI world has to a public scoreboard, within about a week of release.

Scaling Up, Then Scaling Down

TII didn't stop at 40 billion parameters. By September 2023 it had released Falcon 180B, a 180-billion-parameter model pitched as one of the largest openly available LLMs at the time, licensed for both research and commercial use. It was an enormous model to run — the kind of thing that needs serious server hardware, not a laptop.

That turned out to be the high-water mark for TII's "bigger is better" phase. The more interesting shift came afterward, when the team pivoted toward smaller, more efficient models that regular developers could actually run. Falcon 2, released in May 2024, included an 11-billion-parameter model and a vision-language variant, Falcon 2 11B VLM, that could reason about images as well as text. Gulf News reported at the time that TII said the new model outperformed the comparably sized open model from Meta, and that the institute planned to move Falcon's long-term development under a newly created "Falcon Foundation" — explicitly modeled on the Linux Foundation's governance structure, meant to give the open-source community more of a say in where the project goes.

Then came Falcon 3 in December 2024, which doubled down on efficiency rather than size. TII released a family of text models — 1B, 3B, 7B, and 10B parameters — trained on 14 trillion tokens, more than double the 5.5 trillion tokens behind Falcon 2, according to the institute's launch materials. The pitch this time wasn't "biggest open model in the world," it was the opposite: models small enough to run on a laptop or light edge hardware, while still landing at the top of Hugging Face's leaderboard for their size class at launch. TII also folded Falcon Arabic into the lineup, a model specifically tuned for Arabic-language tasks — a deliberate bet that the next wave of useful AI needs to work well in languages Silicon Valley labs often treat as an afterthought.

Why a Government Lab Open-Sources Its AI

It's worth sitting with how unusual TII's approach actually is. Most of the companies racing to build frontier AI — OpenAI, Anthropic, Google DeepMind — keep their most capable models locked behind APIs, charging per token and guarding the weights as trade secrets. TII, funded by an Abu Dhabi government research council, has instead released successive Falcon generations under genuinely open licenses that let developers download the full model weights, run them on their own hardware, fine-tune them for a specific industry, and in many cases use them commercially with no royalty owed back to TII.

The strategic logic isn't mysterious. Open-sourcing a strong model is one of the fastest ways to get global developers building on your technology rather than a rival's, and it buys the UAE a seat at the table in a conversation that's otherwise dominated by a handful of American and Chinese labs. It also lines up with the broader national push — visible in everything from the UAE's sovereign AI investments to its push for homegrown digital infrastructure — to be a producer of core AI technology, not just a buyer of it.

What It Actually Means for Developers

For anyone actually building software, the practical upshot is that Falcon's smaller models are a legitimate option alongside Llama, Mistral, and Qwen when picking an open-weight model to self-host — particularly for Arabic-language products, where TII's models are trained with more local-language data than most Western alternatives. You won't find Falcon powering a mainstream consumer chatbot the way GPT or Gemini do, but inside the open-source AI ecosystem, it's a genuinely respected line of models, not a token national-pride project.

Falcon 3, released in December 2024, is the most recent major version TII has detailed in full; the project's momentum — and the newly formed Falcon Foundation — suggest more releases are still to come. Either way, it's a reminder that some of the more consequential AI research of the last few years hasn't come out of a Bay Area campus, but out of a government lab in Abu Dhabi that decided the smartest move was to give its work away.

Sources: Technology Innovation Institute (tii.ae); Emirates News Agency (WAM); Gulf News technology coverage.