Chuyện gì đã xảy ra
Yandex released AliceAI-Foundation-80B-A3B-Base, an 80 billion- large language model, on Hugging Face under an Apache 2.0 license. The model utilizes a Mixture-of-Experts architecture that activates only 3 billion parameters per token, supports a 262,144-token context window, and was trained on approximately 18 trillion tokens. It is designed for both commercial deployment and academic research, with specific optimizations for Russian-language tasks.
Yandex has made AliceAI-Foundation-80B-A3B-Base available on Hugging Face under an Apache 2.0 license, allowing unrestricted commercial use, modification, and product development without licensing fees. The model is an 80 billion- system built for both commercial deployment and academic research.
The architecture employs a Mixture-of-Experts (MoE) design where only 3 billion parameters are active for any given token. The routing system selects the top 10 experts from a pool of 512, plus one shared expert that remains always active. This design runs across 48 layers using a hybrid of Kimi Delta Attention and gated attention, supporting a context window of up to 262,144 tokens.
The model was trained from scratch on approximately 18 trillion tokens. Yandex benchmarked it against DeepSeek variants, reporting that it outperforms DeepSeek-V4-Flash-Base on two new Russian-language benchmarks introduced alongside the release: WikiWebFacts and HardMultiQA. On coding tasks, the model reportedly holds its own against NVIDIA’s Nemotron-3-Super-120B despite activating far fewer parameters per forward pass.
This release follows a 35 billion- open model released by Yandex just two weeks prior in September 2026. It represents an architectural leap from dense transformer designs to the MoE approach, continuing Yandex's lineage of open-source AI that includes the earlier YaLM 100B model.
Chi tiết nguồn: cryptobriefing.com ↗
Tại sao nó quan trọng
This release provides a high-performance, open-licensed alternative for Russian-language applications, addressing a gap where most open models are predominantly English-trained. The MoE architecture lowers hardware requirements for on-premise deployment, making it more accessible for regulated industries like law and medicine. By outperforming DeepSeek-V4-Flash-Base on new Russian-language benchmarks, it establishes a competitive baseline for sovereign AI efforts in the region.
The model is explicitly framed as a sovereign AI solution, highlighting performance in Russian legal and medical contexts where domain accuracy is critical and existing Western or Chinese models may fall short due to training data imbalances.
For developers working in Russian-language applications, this model fills a significant gap, as most high-quality open models are predominantly English-trained. The ability to use the model without licensing fees lowers barriers to entry for businesses and researchers.
The MoE architecture has practical implications for deployment costs. 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.
Cơ chế tương tác: Nó thực sự hoạt động như thế nào
Khám phá công nghệ cơ bản đằng sau sự phát triển này một cách tương tác.
In AI, what are a model's "parameters"?
Xem gì tiếp theo
Monitor the adoption of the new WikiWebFacts and HardMultiQA benchmarks by the broader AI community as potential standards for evaluating Russian-language models. Watch for the integration of this foundation model into Yandex's consumer-facing Alice assistant for future reasoning and agentic capabilities. Track whether other organizations adopt the model for on-premise deployments in regulated sectors.
The two new benchmarks, WikiWebFacts and HardMultiQA, may become community standards for evaluating Russian-language models, influencing how performance is measured in this specific linguistic domain.
Yandex intends this foundation model to serve as the engine for future reasoning and agentic capabilities in its consumer-facing Alice assistant, suggesting upcoming product updates that leverage this new infrastructure.
Independent verification of the benchmark claims is currently limited to Yandex's own reporting, as no third-party evaluations have been cited in the source material.