German AI startup Aleph Alpha has released a new AI tool enabling customers to customise third-party models, incorporating different languages and industry-specific data.
The company announced last year that it was moving away from training its own LLMs, to focus on developing a “generative AI operating system” to sell to B2B customers — such as large enterprises and local governments — to help them roll out AI in their organisations.
For example, the new technology could help companies build an internal chatbot that recognises different languages and company-specific terminology.
Aleph Alpha’s CEO Jonas Andrulis says he’s seen a lot of demand for the solution. He told Sifted. “I’ve spoken to many countries and enterprises that are trying to fine tune their specific knowledge, technical documents and languages into these language models and it fails.”
Aleph Alpha raised $500m (€460m) at the end of 2023, €110m of which was equity, with the rest in the form of research grants.
What problem is this solving?
Aleph Alpha says existing methods for fine-tuning — the task of making general purpose models more specialised by feeding them extra data — has been difficult to do due to a process known as “tokenisation”.
Tokenisation involves translating language and words into long strings of numbers that allow AI models to analyse them statistically, something the company says is difficult to do if the new data is significantly different from the original training data.
“If new text is sufficiently different from the original training data, it cannot be tokenized efficiently,” Aleph Alpha said in a statement, announcing its new architecture that “requires no tokenizer and therefore has no limited vocabulary.”
“If you want to use Meta’s Llama in German, for example, you have basically double the cost of running a system,” Andrulis tells Sifted.
He says Aleph Alpha is able to add Finnish language proficiency to train a model, and can save 70% of the compute cost compared to Llama.
Meta has done similar work to remove tokenizers: Andrulis says it’s possible that, in the future, all state-of-the-art models will be token-free.
Split opinions
Once the darling of Germany’s AI scene, Aleph Alpha has often been described as the country’s answer to OpenAI — but the company’s move away from exclusively training LLM’s seems to have split opinions about the future of the company.
One investor in Aleph Alpha told Sifted that the company’s decision to develop a generative AI operating system and focus on offering infrastructure-as-a-service makes good business sense for the company, rather than competing with the likes of France’s Mistral to build a global LLM player which requires a lot of capital.
However, they added, the new AI architecture is not “completely unique” and did not “knock (their) socks off.”
Other investors like Dr Andre Retterath from Earlybird VC say helping companies use AI was a natural development in the company’s journey — and it doesn’t need to build the best LLMs to do that.
“If you want to serve B2B clients, the full stack solution that enables customers to integrate different models into their applications doesn’t require the most powerful models, but more specialised, tailored models,” he says.
What’s exciting about Aleph Alpha’s operating system, he adds, is that customers “don’t need to care about any part of the process to bring their domain-specific knowledge into a large language model, integrate it into their existing workflows and start delivering value.”
Earlybird increased its stakes in the company in a secondary transaction, while investors 468 Capital and Lakestar sold their shares.
Andrulis says he’s happy that his company has chosen not to spend “hundreds of millions” on building a European LLM.
“Llama 4 will launch this year, we’ve just seen the launch of DeepSeek which at first glance looks like a phenomenal model. So all the models that we were celebrating six months ago are outdated and are worth zero. So I’m super happy that I haven’t bet on this business model,” he says.
“Just having another state of the art language model alone would not be sufficient for our customers.”
Andrulis is hoping the new architecture announced today will unlock a host of new customers in different geographies, which will help its bottom line.
Aleph Alpha failed to reach its own revenue goals in 2023, when it brought in under €1m after targeting €5.5m. It planned to 20x that revenue in 2024.
Andrulis declined to give specific figures, but says Aleph Alpha’s revenue is “increasing drastically”, with much of it coming from multi-year contracts.
Read the orginal article: https://sifted.eu/articles/aleph-alpha-what-comes-next-news/
