Aleph Alpha takes on OpenAI with GPT-3 rival and 300B model in testing

Ever since ChatGPT took over the internet, Europe has been facing an uncomfortable truth: nearly all the foundation models powering the current AI boom are American or Chinese.

Now, a startup from Heidelberg is trying to change the score. Aleph Alpha just released benchmark results showing that its flagship model can go toe-to-toe with OpenAI's original heavyweight, GPT-3.

Matching GPT-3 with less than half the parameters

The model in question is Luminous-Supreme, a 70-billion-parameter system released in August 2022. Aleph Alpha pitted it against OpenAI's 175-billion-parameter GPT-3 davinci, Meta's 175-billion-parameter OPT, and BigScience's 176-billion-parameter BLOOM.

Using the evaluation framework from open-source collective EleutherAI, the models were tested across classification, common sense reasoning, and natural language inference.

Luminous-Supreme matched GPT-3 davinci across almost every benchmark, and even managed to pull ahead in one category. Meanwhile, Meta's OPT and BigScience's BLOOM trailed both by several percentage points. Because Luminous achieves this with less than half the parameters, it runs with significantly lower compute requirements and better real-world scalability.

Prepping a 300B heavyweight for the next round

There is, however, an obvious catch: OpenAI unveiled GPT-3 back in May 2020. It took more than two years for a European champion to match a milestone that competitors in the US, Israel, and China cleared long ago. In the meantime, OpenAI moved on to instruction-tuned versions like text-davinci-003, and Google rolled out its massive 540-billion-parameter PaLM.

Aleph Alpha is well aware of the moving target. This month, the startup introduced Luminous Supreme Control to match OpenAI's newer offerings.

More importantly, the team is actively testing Luminous-World, a 300-billion-parameter model designed for complex, critical tasks. Scheduled for release later this year, it is aimed directly at competing with OpenAI's upcoming GPT-4.

Founder and CEO Jonas Andrulis is also steering the company toward transparency and explainability, arguing that ethical alignment and auditability are non-negotiable requirements for European enterprises.

Why it matters

For European enterprises and developers, this is the first real sign of digital sovereignty in generative AI. Organizations operating under strict data protection rules now have a credible local alternative that does not require routing proprietary data through US cloud infrastructure.

On the industry level, Aleph Alpha proves that raw parameter scale is not the only metric that counts. Matching a 175B model with 70B parameters underscores how much efficiency can be gained through smarter training and model architecture, giving smaller labs a viable playbook against tech giants.

We think positioning the company around explainability and regulatory compliance is a very smart bet. Competing purely on raw chatbot hype against Silicon Valley is a losing game in Europe, but offering transparent, auditable models solves an immediate problem for banks, healthcare providers, and industrial giants.

The real danger is capital and infrastructure. Europe's funding ecosystem still lags far behind the billions pouring into US labs. Keeping up with OpenAI and Google across multiple generations of frontier models requires sustained supercomputing resources, and if the upcoming 300B model stumbles on execution, the gap could widen right back up.