Yandex open-sources AliceAI-Foundation-80B base model

Yandex released AliceAI-Foundation-80B-A3B-Base, an 80-billion-parameter mixture-of-experts language model that turns on about 3 billion parameters per token, trained from scratch and published under Apache 2.0 on Hugging Face with a long technical report.

SOFTWARE desk — a Russian tech company is releasing a from-scratch open mixture-of-experts base, with tool-use practice already in the training data, which widens the open-weight race beyond China and the big US labs.

What it is, still those pages. A language model writes the next piece of text. This one is a mixture-of-experts model, or MoE: a large set of specialist blocks, with only a few turned on for each token. A token is a small chunk of text, roughly a word piece. Yandex says the model has 80 billion parameters in total and 3 billion active per token. A parameter is one learned number. Eighty billion is the whole warehouse. Three billion is the slice that actually runs for each word piece. The rest stay in the model, but they are not all used on every token. File the 80 billion / 3 billion split as Yandex’s, on both the Habr post and the model card. This desk did not count the file.

The shape, still Yandex. Both pages call it a hybrid design with MoE layers. Habr says it is a transformer — the standard modern language-model layout — with 48 blocks and one extra multi-token prediction layer, a small add-on that tries to guess more than one next piece at a time. Each MoE layer has 512 routed experts, keeps the top 10, and adds one shared expert that always runs. Of the 48 attention blocks, Habr says 36 use Kimi Delta Attention, a cheaper way of looking at earlier text, and 12 use full attention, one full block after every three cheaper ones. The Hugging Face card prints the same sandwich and names those 12 blocks Gated Attention. File both wordings. Do not merge them into a third name. The card lists a context length of 262,144 tokens. Context length is how much text the model can hold at once. 262,144 tokens is on the order of a long book, or a large code folder, in one sitting. Habr’s training table includes a stage at that length. The last training stage, Habr says, dropped the context to 128,000 tokens to save compute, because the useful reasoning examples were shorter. File 262,144 as the card’s listed context, and the 128,000-token last stage as Habr’s. This desk did not time a long prompt.

Trained from scratch, Habr is explicit. The model went through a full training cycle at Yandex from zero, and Yandex says it did not use weights from other open models. The December technical report on the earlier Alice AI LLM started from Qwen3-235B-A22B weights. Qwen is Alibaba’s open model family. This run did not start there. Habr says that earlier Alice AI LLM 235B, an October 2025 release, was closed — the weights were not public — and that the new model comes out ahead of it on factual knowledge, math, programming, and long context, with almost three times fewer total parameters and about seven times fewer active parameters. On one Russian fact quiz, WikiWebFacts, the gap versus that older base is 86.5 to 86.2. That is a narrow edge on that set, and it is still Yandex’s own measurement. File the from-scratch line, the Qwen starting point of the older model, and that in-house comparison as Yandex’s. A smaller model ahead of the lab’s own older model on the lab’s own tests is not a win over every other lab.

Why Yandex says it built this. The Habr post calls the release another step toward a unified reasoning model, the base on which Alice AI’s agent features will grow. Alice is Yandex’s assistant. An agent, here, means software that can take a series of steps — look something up, call a tool, write code — not only answer one question. Habr says training on hard reasoning was not enough for those skills. Adding tool interactions already during pretraining improved agent benchmarks after a later post-training step. Post-training is the extra tuning after the base model exists. A benchmark is a fixed test set. Habr also says the path from start to this release took about half a year, because the team had already learned the earlier run. In that time they rebuilt the training set, tried architectures, picked settings, and added data for reasoning and for those tool interactions. Key choices were checked in separate from-scratch trainings of 2 trillion tokens each. File that purpose, that half-year line, and those data notes as Habr’s. The released weights are the base model, not the finished Alice agent.

The scoreboard is Yandex’s own. Habr says that on pretrain measurements the new model comes out ahead of larger open bases, including DeepSeek-V4-Flash-Base and Nemotron-3-Super, on many tasks from factual and expert questions through math and code, and that on the hard IMO AnswerBench and LiveCodeBench it leads the open pretrains they compared. The Hugging Face card is softer on that point: it says math, programming, and other reasoning tasks land at the level of larger open models, and that the model is especially strong on factual knowledge in Russian. The comparison set on the Habr page is Qwen3.5-35B-A3B-Base, GLM-4.5-Air-Base (106 billion total, 12 billion active), Nemotron-3-Super-120B-A12B-Base, and DeepSeek-V4-Flash-Base (284 billion total, 13 billion active), plus Yandex’s own older 235B. Yandex notes its model has fewer total and active parameters than every model in that set except Qwen, which has fewer total parameters and the same 3 billion active. Habr says the model takes first place on 8 of 10 benchmarks in the factual, education, and expert group. Both pages say the first score table was run on Yandex’s internal setup, in vLLM — software for running the model — at temperature 0, which means the model picks the most likely next token instead of rolling the dice. File every score as Yandex’s measurement, not as a number this desk reproduced.

A few of those Yandex figures, so the claim has a scale. They are from the temperature-0 table. On WikiWebFacts, a Russian factual quiz Yandex built, the model scores 86.5, against 83.2 for DeepSeek-V4-Flash-Base, 72.8 for Nemotron-3-Super, and 62.4 for the Qwen base. On HardMultiQA, the harder Russian factual set, it scores 67.9, against 65.4 and 54.5 for those two larger bases. On MATH-500, a math set, it scores 91.1, against 84.8 for Nemotron and 80.7 for DeepSeek. On a LiveCodeBench coding check (versions 5 and 6, with a reasoning example in the prompt, pass@1), it scores 50.5, just ahead of Qwen and Nemotron at 50.4 and ahead of DeepSeek at 38.1. Pass@1 means the first try counted. On English TriviaQA the picture flips: Yandex reports 79.0 for its model, against 89.8 for Nemotron and 89.4 for DeepSeek. Strong on the Russian sets is not the same as leading English trivia. Habr’s “leads on IMO AnswerBench” line is about a second, harder table, where the model gets many tries and a task counts if any try is right. Habr says DeepSeek’s base stayed near zero on that zero-shot setup, so DeepSeek is missing from that second table. File that exclusion as Yandex’s, not as a DeepSeek claim.

What shipped with the weights. Habr says a long technical report goes out with the model, covering how the data was prepared, how the tests were scored, how training was kept stable, and scaling laws — the rules of thumb for how quality changes when you add data or size. Alongside the weights, Yandex is opening two factual benchmarks aimed at Russian-language context, WikiWebFacts and HardMultiQA, plus the instructions for scoring them. Habr says publishing them lets other people reproduce the report’s numbers. The Hugging Face card repeats that those two benchmarks and their scoring instructions ship with the model. File the report and the two open tests as Yandex’s. This desk did not grade a copy.

What this release is not. The Hugging Face card says the model is pretrained and is provided as-is, with no extra post-training or alignment step. Alignment is the tuning that makes a model follow the habits of a user-facing assistant. The card says it is meant for research, experiments, and further work. It is not a finished product for a consumer service. Habr’s agent scores were checked after a supervised fine-tune — a shared extra training step on agent examples — not on the raw base alone. Do not treat this card as a match against Claude or GPT. Those names are not in Yandex’s comparison, and this is a base model, not a finished assistant. Do not invent a Western newsroom confirmation. None is on this card.

Plain English for the rest of the card: MoE / mixture-of-experts = a model with many specialist blocks, only a few active per token. parameter = one learned number inside the model. token = a small chunk of text. context length = how much text the model can hold at once. Apache 2.0 = the open-source license on these weights. base model / pretrain = the model after the big first training, before assistant polish. from scratch = training started from random numbers, not from another lab’s weights. agent = software that can take a series of steps and use tools. benchmark = a fixed test set. This filing is the open base-model release. It is not a finished Alice assistant, and it is not a score this desk reran.

PRIMARY here: Yandex’s 21 Sep 2026 Habr company post and the Hugging Face model card for yandex/AliceAI-Foundation-80B-A3B-Base — Tier A PRIMARY, the original record. There is no independent Western wire on this card. The Apache 2.0 opening, the from-scratch line, the older model’s Qwen3-235B-A22B start, the unified-reasoning-model and Alice agent framing, the half-year line, the 2-trillion-token check runs, the comparison set, the score lines, the technical report, and the WikiWebFacts and HardMultiQA release are Yandex-attributed on Habr. The license field, the 80-billion / 3-billion split, the hybrid layer pattern, the 262,144-token context, the Russian-facts emphasis, the “at the level of larger open models” line, and the no-post-training disclaimer are also on the Hugging Face card. NOT claimed: an independent rerun of the scores, a win over Claude or GPT, that the base model is a finished assistant, that agent scores were measured on the raw base, a Western newsroom confirmation, that this desk downloaded the weights, a stock tip, or investment advice. Distinct from the already-filed stepfun-step-5-preview, zai-zcode-open-source, and spacexai-grok-4-7.