Yandex has released the weights of AliceAI-Foundation-80B-A3B-Base under the Apache 2.0 license
· Source: original
Yandex has open-sourced the weights of AliceAI-Foundation-80B-A3B-Base under Apache 2.0
80B parameters, of which 3B are active. Trained from scratch with a full cycle at Yandex, without weights from third-party open-source models.
For practical use, this is a ready-made base 🧠 Apache 2.0 — a license without restrictions, and out of the 80B, only 3B active parameters run at inference. After that, it's just fine-tuning for your task.
In pretrain benchmarks, the model outperforms DeepSeek-V4-Flash-Base and Nemotron-3-Super on factual, expert, and mathematical tasks, as well as on code. On the challenging IMO AnswerBench and LiveCodeBench, it leads among the compared open pretrains.
The new model surpasses the closed Alice AI LLM 235B in factual knowledge, mathematics, programming, and long-context handling. And this with almost three times fewer total parameters and roughly seven times fewer active ones.
The positioning is a base for a unified reasoning model (URM), on which the agentic capabilities of Alice AI will be developed. This is specifically a Base pretrain, not a final reasoning model.
The publication is dated September 21, 2026 — the weights were posted in Yandex's blog on Habr. Earlier, Yandex colleagues had already open-sourced the Alice AI Search pretrain, also trained without weights from open-source models
🤖 Interested in AI agents and automation?
Prompts for building AI agents and automations — read up on the topic:
🔗 The entire prompt library · "AI Agents" category
A ready-made product on the topic: Prompts for Programmers — grab it and apply it right away.