Naive AI, founded in February 2026 by a Tsinghua professor, reaches a valuation of around $1.42 billion after raising $400 million across three rounds. The team numbers fewer than one hundred people. Its method: it does not pre-train any model from scratch. It takes open-weight models and reworks them through phases of intermediate training, fine-tuning and reinforcement learning.
We have been describing a shift in value for months. Here is the clearest case we have seen, and it comes down to one figure: seven months, one hundred people, $1.42 billion.
What this method assumes
Pre-training a model from scratch costs hundreds of millions and requires skills that very few teams possess. That is what reserved model production for a handful of players, as we explained in our article on foundation models.
Starting from an open model changes the equation entirely. You inherit general capabilities for free — command of language, knowledge, basic reasoning — and you invest only in what differentiates.
We have seen this logic at work several times this year: a legal publisher building its model on an open base for forty million, a Saudi lab producing a 428-billion Arabic model on Chinese weights.
What is new here is that specialisation is no longer a means: it is the product.
Why investors are funding this
The obvious objection is that a company building on someone else's work has no asset of its own. It is weaker than it appears, for three reasons.
Specialisation know-how is rare. Taking a general-purpose model and making it excellent in one domain requires skills few have mastered: building the right datasets, designing the evaluations, balancing the phases without degrading what already worked.
Speed is an asset. A team of one hundred people delivering in a few weeks what a large lab takes months to produce occupies a real slot.
Supply is assured. This is the decisive point. Two years ago, depending on open models was a gamble. Today, several labs publish leading models, and competition between them guarantees more will follow.
This company occupies a position we see emerging everywhere: that of the integrator. It does not produce the raw material, it does not sell to the end customer, it transforms. That is exactly the role historically played by equipment suppliers in the automotive industry or by contract manufacturers in pharmaceuticals. Their appearance in a sector is an indicator of maturity: it signals that the underlying layer has become commoditised to the point where you can build on it without mastering it.
The fragility of the model
The caveat has to be stated, because it is real.
A company that builds on open models depends on their continued existence. If the labs that publish them were to stop, or to move to restrictive licences, the whole edifice would be under threat. This is the point we set out in our article on licences.
And specialisation itself becomes commoditised. What requires an expert team today may be a standard feature in eighteen months.
What to take away
This valuation is a signal, and it concerns far more than this one company.
For two years, the question was who would have the best model. This company says something else: the question is who will be best at using the ones that exist.
This is exactly the conclusion we drew for Europe in our article on the need to be the best, and for companies in our framework for building versus buying. It now holds for an investor too: you can value at 1.4 billion a team of one hundred people that owns no model.