Superpower Daily

Superpower Daily

Your reading space

Sign in
New York Sets November Registration for Frontier AI DevelopersAI Panel Publishes Warning on Losing Control of Agents After Op...xAI Releases Grok 4.7 With Longer Reasoning at Grok 4.6 PricesMicrosoft Research Open-Sources RetroChimera for Chemical Synth...Google Starts Googlebook Preorders, Putting Android Phone Apps...WIRED Review Finds Memory Limits Apple’s M6 Mac Mini for AI Age...New York Sets November Registration for Frontier AI DevelopersAI Panel Publishes Warning on Losing Control of Agents After Op...xAI Releases Grok 4.7 With Longer Reasoning at Grok 4.6 PricesMicrosoft Research Open-Sources RetroChimera for Chemical Synth...Google Starts Googlebook Preorders, Putting Android Phone Apps...WIRED Review Finds Memory Limits Apple’s M6 Mac Mini for AI Age...
Swipe

Yandex Releases Alice AI Foundation, an 80B Model for Commercial Use

The downloadable base model uses only 3 billion parameters per token, comes with Russian-language factual benchmarks, and sits ahead of Yandex’s planned reasoning and agent work for Alice AI.

By 2 min read
Yandex Releases Alice AI Foundation, an 80B Model for Commercial Use
Yandex Releases Alice AI Foundation, an 80B Model for Commercial Use

Listen to this story

The audio brief

About 1:34
0:001:34
Read transcript
Yandex has released Alice AI Foundation, an 80-billion-parameter base model that developers can download and use commercially or for research. Its architecture is a mixture of experts: the model contains 80 billion parameters overall, but activates about 3 billion for each token, routing different text through specialized parts instead of running the entire system every time. That gives developers access to a large model with a different compute profile from a dense 80-billion-parameter system. The model was trained from scratch and supports a context window of up to 262,144 tokens—enough to handle unusually long documents, which is one of the clearest practical reasons to pay attention to this release. Yandex has also published WikiWebFacts and HardMultiQA, two Russian-language factual-knowledge benchmarks, along with evaluation protocols. The company says Alice AI Foundation is competitive with larger open-source models in mathematics, coding, and other reasoning tasks, and especially strong on Russian factual knowledge. But those comparisons come from Yandex’s own infrastructure and inference settings, so they are not an independent benchmark assessment. Strategically, the model is intended to underpin future reasoning and agent capabilities in Alice AI. Yandex calls it sovereign, but that status is not yet official. Russia’s registration process begins on March 1, 2027, and sovereign models will need entirely domestic components. The key constraint is that the label remains undecided until that process starts.

Story brief

3 key points

Yandex’s release gives developers a commercially usable 80B-parameter mixture-of-experts base model with a 262,144-token context window, making it relevant for long-document and Russian-language applications. Only 3B parameters activate per token, creating a different compute profile from a dense 80B model. The package includes two Russian factual-knowledge benchmarks and evaluation protocols. Planned use in Alice’s...

  1. 01

    Alice AI Foundation contains 80B total parameters but activates 3B per token through a mixture-of-experts architecture.

  2. 02

    The 262,144-token context limit supports unusually long inputs for document-heavy applications.

  3. 03

    WikiWebFacts and HardMultiQA target Russian factual knowledge and include published evaluation protocols.

Developers can now download and use Yandex’s new Alice AI Foundation model for commercial and research work. The release puts an 80-billion-parameter base model into public reach, while activating 3 billion parameters for each token it processes.

Alice AI Foundation uses a mixture-of-experts design, meaning it has many specialized components but routes each piece of text through only part of the full model. Yandex says the model has 80 billion parameters in total, with 3 billion activated for each token. The company trained it from scratch and says its context window can reach 262,144 tokens, allowing it to take in lengthy inputs.

Yandex also released WikiWebFacts and HardMultiQA, two factual benchmarks focused on Russian-language contexts, along with their evaluation protocols. Its model card says Alice AI Foundation performs on par with larger open-source models on mathematics, coding and other reasoning tasks, and is particularly strong on Russian factual knowledge.

Those comparisons are company-run results. The repository says Yandex used its own internal evaluation infrastructure and specific inference settings for the published tables. That makes the release more useful to developers as a set of weights and reproducible evaluation materials, but it does not turn the reported rankings into an independent assessment.

Yandex plans to use the model as a foundation for future reasoning-model and agent capabilities in its Alice AI product. It also describes Alice AI Foundation as sovereign, but formal status has not yet been assigned: Russia’s framework for sovereign models is still being developed, and registration and status assignments are scheduled to begin March 1, 2027.

The new rules set a higher bar for sovereign models than for national models, including a requirement to rely entirely on domestic components. For now, the concrete change is the open release itself: a public base model, a long context limit and an evaluation package aimed at Russian-language knowledge. Whether it receives the sovereign label will depend on a process that has yet to start.

Sources

  1. huggingface.coREADME_en.md · yandex/AliceAI-Foundation-80B-A3B-Base at main
  2. cryptobriefing.comYandex releases 80B-parameter AI model under open license for commercial use

Loading discussion...

YOUR READING SPACE

Notifications

Yandex Releases Alice AI Foundation, an 80B Model for Commercial Use | Superpower Daily