Official English account for Yandex. We build AI and put it to work at scale across Search, Maps, Cloud and devices. Research, open source and product news.
34 global offices
Joined October 2012
- Tweets1.4K
- Following17
- Followers35.8K
- Likes129
Five Yandex scientists and educators rank in the global top 1–2% in their fields, per @Stanford & @Scopus . A sixth made the career-long top 1%.
elsevier.digitalcommonsdata.…
Their work includes #CatBoost, a widely used gradient boosting algorithm, and AQLM, an LLM quantization tool. 🧵
AQLM tackles a major challenge in running large language models: memory consumption.
It compresses model weights to just 2–3 bits per parameter, dramatically reducing memory requirements. It's open source and supported by Hugging Face Transformers.
huggingface.co/docs/transfor…
RainNet uses neural networks and radar data to forecast rainfall. Further research targets more accurate heavy-rain predictions — critical for anticipating extreme weather.
github.com/hydrogo/rainnet
From LLMs to weather, this work shows how research can make a difference.
Fresh recommendations need fresh data — but real-time processing at scale is hard.
That’s why we built YTsaurus Flow. Now open source, it handles 100+ GB/s, keeps state between events and provides exactly-once by default.
Apache 2.0 👉github.com/ytsaurus/ytsaurus…
The trick is in the infrastructure: managing state, balancing load and recovering from failures without losing or duplicating data.
On our blog, we break down how YTsaurus Flow handles those challenges — and why we built our own engine.
medium.com/yandex/open-sourc…
Our researchers at @YandexResearch are exploring new ways for LLMs to operate as agents. Meet AsyncLLM: a framework that lets pretrained models observe, reason and act concurrently — with no additional training.
Paper, code and demos 👇
LLMs are general asynchronous agents
We're releasing AsyncLLM - a framework for pretrained LLMs to observe, reason and act concurrently, with no additional training
Paper: huggingface.co/papers/2609.3…
Code: github.com/dvmazur/async_llm
Website with Demos: dvmazur.github.io/async_llm
🧵...
Sona reads a user’s history once and uses it to both generate and rank candidates, avoiding repeated computation across stages.
It builds on a decade of our recommendation research, including ARGUS and Gryphon. Our open dataset Yambda, with nearly 5B interactions, has also been cited by @Meta and Kuaishou researchers.
Teams at ByteDance, @TencentGlobal, @Spotify and @PinterestEng are advancing the same broader shift toward unified recommendation models.
Explore Sona’s architecture and results: arxiv.org/abs/2608.11015
A look behind the scenes: Running our 8B data filter over the whole corpus would have cost more than 200,000 GPU-hours. So we built a multi-stage pipeline that keeps about 95% of the useful documents.
Full report on arXiv soon.
Learn more about training AliceAI-Foundation-80B-A3B-Base here: medium.com/yandex/aliceai-fo…
⚡80B parameters, only 3B active per token.
We've released the weights of our Alice AI Foundation 80B-A3B Base. It's open for anyone to build on. 256K context, Apache 2.0.
Explore it on Hugging Face: huggingface.co/yandex/AliceA…
⚡80B parameters, only 3B active per token.
We've released the weights of our Alice AI Foundation 80B-A3B Base. It's open for anyone to build on. 256K context, Apache 2.0.
Explore it on Hugging Face: huggingface.co/yandex/AliceA…
📌 Alice AI Foundation 80B-A3B Base — EN README on Hugging Face: huggingface.co/yandex/AliceA…
This write-up from @HuggingModels went out two days before our own announcement. Thanks for the head start 🙌
The model comes with no post-training, so it's yours to fine-tune.
“Thinking longer” for LLMs doesn’t have to mean “responding later.”
Rethinking the inference runtime can make pretrained models more interactive without changing their weights.
Proud to see Yandex Research pushing new ideas in interactive AI👏
New blogpost from our researchers 🧊
We discuss our recent work on how sharing and scheduling KV-cache state lets pretrained LLMs observe, reason, and act concurrently without additional training
research.yandex.com/blog/the…
AI is moving beyond one-model-fits-all systems. With Expert mode, Alice AI chooses the model best suited to each task.
Quick questions stay quick, while complex tasks get more resources. A shared harness keeps tools and context working across models.
medium.com/yandex/alice-ai-n…
Behind the scenes, we're digitizing:
📜 roughly 2 million pages of manuscripts and drafts
🗄️ 20,000+ photos from Tolstoy's personal archive, already filterable by genre
In the next phase, neural networks will help identify people across the photos and link related images together.
This is a joint project with the Leo Tolstoy State Museum-Reserve, the institution responsible for preserving Tolstoy's archives.
This is just the start. Every project we build together in the coming years will appear in one place.
Follow along ⬇️
yandex.com/tolstoy