Yandex open-sources an 80B model it says it trained from scratch

Yandex released AliceAI-Foundation-80B-A3B-Base, an 80-billion-parameter mixture-of-experts base model, under Apache 2.0, describing it as trained from scratch.

In this article

What is inside the model
Benchmarks are the company's own
Availability

Yandex open-sourced AliceAI-Foundation-80B-A3B-Base under an Apache 2.0 license around 21 September 2026. The company describes it as a base model trained entirely from scratch — not initialized from Qwen, Llama or another existing model's weights — released as an experimental testbed for the architecture behind Yandex's planned unified reasoning model for its Alice assistant.

A network switch in a server rack with blue and green Ethernet cables plugged in. — Photo: Helpameout (CC BY-SA 3.0), Wikimedia Commons.

What is inside the model

AliceAI-Foundation-80B-A3B-Base is a hybrid mixture-of-experts model with about 80 billion total parameters and roughly 3 billion active per token, using 512 experts with a top-10-plus-shared-expert routing scheme and a mix of attention layer types. Context length runs up to 262,144 tokens. It is explicitly a base model — not instruction-tuned and not a chatbot — so it needs a fine-tuning step before it behaves like an assistant.

Benchmarks are the company's own

Yandex benchmarks the model against Qwen3.5, Nemotron-3 and DeepSeek-V4-Flash-Base and reports it matching larger open models on coding and reasoning tasks while using less inference compute, thanks to the small share of active parameters per token. These comparisons come from Yandex's own release materials rather than an independent evaluation, and should be read as the company's account of where its model lands, not a neutral ranking.

Availability

The roughly 163 GB full-precision model is published on Hugging Face under the Apache 2.0 license, alongside a technical report and the WikiWebFacts and HardMultiQA evaluation sets used in Yandex's own testing. The community has already produced quantized GGUF, MLX and NVFP4 variants for cheaper local inference.

For teams evaluating open-weight base models, the Apache 2.0 license and the from-scratch training claim are the parts worth checking directly against the model card; the comparative benchmarks are marketing until someone outside Yandex reruns them.

Sources: Yandex — investor relations press release, Hugging Face — yandex/AliceAI-Foundation-80B-A3B-Base.