After six years focusing on closed-source technologies, OpenAI has released two ‘open’ AI models in a move set to heighten the competition for Europe’s homegrown open-source AI champion Mistral.
On Tuesday the US tech giant unveiled two models that are ‘open weights’, meaning their training parameters are free to download, enabling developers to scrutinise and fine-tune the model for their own applications. OpenAI said the models have similar capabilities to some of its industry-leading closed-source models.
It brings it close to Paris-based Mistral, which launched in 2023 to build an open-source competitor to OpenAI. With over $1bn raised in the past two years and a $6bn valuation, Mistral is widely considered Europe’s best shot at competing in the global AI race.
The French startup has faced mounting competition in the open-source space from US tech giants like Meta, and more recently from a host of Chinese players such as DeepSeek, Qwen and Yi. OpenAI entering the space puts Mistral under even higher pressure to convince customers.
“It’s a great move from OpenAI,” says Tariq Krim, a French tech entrepreneur and founder of tech think tank Cybernetica. “Lots of people in Europe will adopt the technology […] It’s a serious competition for Mistral.”
Mistral did not reply to a request for comment.
Competing for enterprises
From its early days Mistral has positioned itself as a challenger to OpenAI thanks to its open-source approach, which it says promotes European values of transparency and data privacy.
With this play the startup has largely targeted enterprise and government customers, which can host Mistral’s models on their own premises to carry out business applications, ensuring better control over the technology and data. It has found particular traction in sensitive sectors, pursuing defence contracts with governments across Europe and Asia.
Mistral also has a consumer-facing product called Le Chat, which is similar to OpenAI’s ChatGPT, the product the US tech company has mostly been known for.
OpenAI’s latest release suggests it’s making a stronger play for enterprise customers. The larger and more capable of the two models unveiled by the company, dubbed gpt-oss-120b, specifically targets enterprises and governments.
OpenAI said it has been working with partners like French telecoms operator Orange and US data storage platform Snowflake to trial real-world applications.
Krim says there’s a strong chance European customers will choose OpenAI’s open-source models for their critical systems.
“The priority will be to pick the best system,” he says. “And an American path seems safer than a Chinese model.”
Recent months have seen Mistral falling behind competitors like OpenAI, Anthropic and Qwen on common benchmarks for model performance, such as the MMLU leaderboard, which compares the multilingual capabilities of large language models (LLMs).
Sell, sell, sell
Yann Lechelle, the cofounder of Paris-based open-source startup Probabl, says OpenAI’s move means foundational model providers like Mistral can no longer rely on their models as a differentiator.
“There is almost no moat left in the models themselves,” says Lechelle. “Everything is about services and distribution.
“From now on to exist you need to sell, sell, sell.”
One way to do this, he says, is to follow the ‘forward-deployed engineer’ concept developed by Palantir, in which software engineers embed directly with customers to deploy the technology, and to adapt and iterate in real-time based on feedback from users.
It’s a strategy Mistral is seemingly already adopting: earlier this year the startup inked a €100m multi-year deal with French shipping giant CMA CGM and dispatched a dedicated team to the corporate’s HQ in Marseille in the south of France.
Mistral’s sales are reportedly on track to surpass $100m a year for the first time in 2025 — although this is a fraction of OpenAI’s figures, which was reported to have hit $12bn in annual recurring revenues (ARR) in June.
The French startup is also pursuing different strategies: earlier this year it launched Mistral Compute, an Nvidia-backed infrastructure platform to provide cloud services and enable European enterprises to avoid relying on US providers for AI workloads.
Mistral is also reportedly in talks to raise $1bn at a $10bn valuation to further the commercial rollout of Le Chat and improve its LLMs.
Mass adoption
For Maxime Corbani, an investor at Runa Capital, a VC firm that focuses on open-source startups, OpenAI’s real win lies in the second model it ‘opened’ this week. Dubbed gpt-oss-20b, the tool can run on a consumer laptop with only 16GB of memory.
“The intent is pretty clear […] they want to appeal to developers playing on device,” says Corbani. “They are aware things are going to be running on-device, and they’re not going to miss out.”
Corbani says the release is primarily a go-to-market move from OpenAI intended to ensure mass adoption, particularly among the vast community of open-source developers. “They want to get adoption very early on from people hacking around and starting to prototype,” he says.
“It’s clearly a way to confirm and reinforce their go-to-market and leadership, and not let others take the lead or engage with a potentially large niche they don’t serve.”
Read the orginal article: https://sifted.eu/articles/openai-open-source-competition-mistral/