Yandex has released the weights for Alice AI Foundation, an 80bn-parameter language model that the company says is available for commercial and research use. It is a substantial open-model launch, but also a revealing bid to build a more self-sufficient Russian AI stack.
Watch Desk analysis
What happened
According to Crypto Briefing, citing Vedomosti, Yandex trained the model on 18 trillion tokens. Only around 3bn parameters are active at any one time, suggesting a mixture-of-experts design intended to deliver large-model capacity without activating the full system for every request.
Yandex plans to use Alice AI Foundation as the base for a future reasoning model and agent capabilities for Alice AI. The company describes it as a sovereign model, although Russia’s legal framework for that status is still being developed. Formal registration is due to begin on 1 March 2027.
The model’s development cost has not been disclosed. Industry estimates cited by Vedomosti range from roughly 150m to 4bn roubles, depending on which stages and expenses are counted. That is a very wide bracket, and a useful reminder that “built from scratch” is not an invoice.
Why it matters
The commercial-use licence could make the model relevant to developers who need Russian-language capability or want to run an open model without handing their work to a foreign provider. The 18-trillion-token training claim and the sparse architecture also place Alice AI Foundation in the increasingly crowded contest to make large models cheaper and more practical to operate.
The political context matters too. Russia’s proposed sovereign-model rules are expected to require entirely domestic components and impose stricter conditions than the “national model” category. Yandex is therefore releasing software into a market where technical ambition and state-defined self-reliance are becoming difficult to separate.
Our read
This earns more than a launch notice because it combines a meaningful open model with a clear infrastructure and policy consequence. Developers should treat the release as an option to evaluate, not as proof of performance: the supplied account does not provide independent benchmarks, model-card detail or evidence that the promised reasoning and agent systems exist yet.
The interesting test is whether Yandex can turn a large, locally controlled foundation model into a usable developer platform. Open weights are the invitation. Adoption, documentation and reproducible results are the reply.
What to watch
- Independent evaluations:
whether researchers publish comparable Russian-language and general-purpose benchmarks. - Developer access:
how complete the documentation, tooling and licence terms are in practice. - Sovereign-model rules:
whether the March 2027 registration process changes the project’s obligations. - The next model:
whether Yandex delivers the promised reasoning and agent capabilities.
Discussion spark: Should open models built for national technological independence be judged mainly on technical performance, or does domestic control of the stack deserve equal weight?
Sources and evidence
- Yandex releases 80B-parameter AI model under open license for commercial use – Crypto Briefing (21 September 2026, 15:50 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.