{"slug": "yandex-releases-80b-parameter-ai-model-under-open-license-for-commercial-use", "title": "Yandex releases 80B-parameter AI model under open license for commercial use", "summary": "Yandex released AliceAI-Foundation-80B-A3B-Base, an 80 billion-parameter large language model, on Hugging Face under an Apache 2.0 license for commercial and academic use. The Mixture-of-Experts model activates 3 billion parameters per token via top-10 routing from 512 experts plus one shared expert, was trained from scratch on approximately 18 trillion tokens, and supports a 262,144-token context window. Yandex says the model outperforms DeepSeek-V4-Flash-Base on its two new Russian-language benchmarks, WikiWebFacts and HardMultiQA, and holds its own against NVIDIA's Nemotron-3-Super-120B on coding tasks.", "body_md": "Photo: Tima Miroshnichenko / Pexels\n\n# Yandex releases 80B-parameter AI model under open license for commercial use\n\nRussia's tech giant open-sources a massive language model trained on 18 trillion tokens, with a design that punches above its weight on Russian-language tasks\n\nYandex has released a large language model called AliceAI-Foundation-80B-A3B-Base, an 80 billion-parameter system built for both commercial deployment and academic research. The model is now available on Hugging Face under an Apache 2.0 license, meaning anyone can use it, modify it, or build a product on top of it without paying Yandex a licensing fee.\n\n## A big model that runs lean\n\nThe headline number is 80 billion parameters, but the more interesting number is 3 billion. That is how many parameters the model actually activates for any given token, thanks to a Mixture-of-Experts architecture. The routing system selects the top 10 experts per token from a pool of 512, plus one shared expert that is always active.\n\nThe model was trained from scratch on approximately 18 trillion tokens and supports a context window of up to 262,144 tokens. Architecture-wise, it runs across 48 layers using a hybrid of Kimi Delta Attention and gated attention.\n\nYandex benchmarked the model against DeepSeek variants and found that AliceAI-Foundation-80B-A3B-Base outperforms DeepSeek-V4-Flash-Base on two new Russian-language benchmarks the company introduced alongside the release: WikiWebFacts and HardMultiQA. On coding tasks, the model holds its own against [NVIDIA](https://cryptobriefing.com/markets/nvidia/)’s Nemotron-3-Super-120B despite activating far fewer parameters per forward pass.\n\n## Why Russia, why now\n\nThe model is explicitly framed as a sovereign AI solution. Yandex highlights its performance in Russian legal and medical contexts specifically, areas where domain accuracy matters and where existing Western or Chinese models often fall short due to training data imbalances.\n\n### AI, tech, and the markets they move—in one daily briefing.\n\nDaily. Free. Join 34,000+ readers across crypto, finance, and policy.\n\nThe company released a 35 billion-parameter open model just two weeks before this one, also in September 2026. Yandex’s history in open-source AI stretches back to YaLM 100B, one of the earlier large Russian-language models released publicly. AliceAI-Foundation-80B fits into that lineage but represents a meaningful architectural leap, moving from dense transformer designs toward the MoE approach.\n\nThe Alice name ties back to Yandex’s consumer-facing AI assistant. Yandex intends this foundation model as the engine for future reasoning and agentic capabilities.\n\n## What this means for the open-source AI landscape\n\nFor developers working in Russian-language applications, most high-quality open models are predominantly English-trained, with Russian as an afterthought. A model that outperforms established competitors on purpose-built Russian-language benchmarks fills a real gap.\n\nThe MoE architecture also has practical implications for deployment. Because only 3 billion of the 80 billion parameters are active at any given moment, businesses can run the model on their own infrastructure more affordably than a fully dense 80B model would require, lowering the hardware threshold for on-premise deployment in regulated industries like law and medicine.\n\nThe two new benchmarks Yandex introduced, WikiWebFacts and HardMultiQA, are worth watching as potential community standards for evaluating Russian-language models.\n\n**Disclosure:** This article was edited by Editorial Team. For more information on how we create and review content, see our\n\n[Editorial Policy](https://cryptobriefing.com/editorial-policy/).", "url": "https://wpnews.pro/news/yandex-releases-80b-parameter-ai-model-under-open-license-for-commercial-use", "canonical_source": "https://cryptobriefing.com/yandex-80b-parameter-ai-model-commercial-release/", "published_at": "2026-09-21 15:50:18+00:00", "updated_at": "2026-09-21 15:52:29.267065+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-products", "ai-startups", "natural-language-processing"], "entities": ["Yandex", "AliceAI-Foundation-80B-A3B-Base", "Hugging Face", "DeepSeek", "DeepSeek-V4-Flash-Base", "NVIDIA", "Nemotron-3-Super-120B", "YaLM 100B"], "alternates": {"html": "https://wpnews.pro/news/yandex-releases-80b-parameter-ai-model-under-open-license-for-commercial-use", "markdown": "https://wpnews.pro/news/yandex-releases-80b-parameter-ai-model-under-open-license-for-commercial-use.md", "text": "https://wpnews.pro/news/yandex-releases-80b-parameter-ai-model-under-open-license-for-commercial-use.txt", "jsonld": "https://wpnews.pro/news/yandex-releases-80b-parameter-ai-model-under-open-license-for-commercial-use.jsonld"}}