---
type: "post"
id: "deepvk-released-user2-base-a-149m-russian-sentence-70f5d3"
slug: "deepvk-released-user2-base-a-149m-russian-sentence-70f5d3"
title: "DeepVK released USER2-base , a 149M Russian sentence encoder with 8,192-token context."
description: "DeepVK has released USER2-base, a 149-million-parameter Russian sentence encoder with an 8,192-token context window."
retrieval_nugget: "DeepVK has released USER2-base, a 149-million-parameter Russian sentence encoder with an 8,192-token context window."
published_at: "2026-10-27"
updated_at: "2026-10-27"
record_date: "2026-10-27"
date_kind: "published_at"
topics: ["models","open-models"]
entities: ["DeepVK","USER2-base","RuModernBERT-base","MTEB-rus","MultiLongDocRetrieval","Sentence Transformers"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "DeepVK released USER2-base, a 149M-parameter Russian sentence encoder with 8,192-token context; no prior document covers this release."
prior_refs: []
source_url: "https://huggingface.co/deepvk/USER2-base"
source_urls: ["https://huggingface.co/deepvk/USER2-base"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:deepvk-released-user2-base-a-149m-russian-sentence-70f5d3"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/deepvk-released-user2-base-a-149m-russian-sentence-70f5d3/","markdown":"https://newruntime.com/posts/deepvk-released-user2-base-a-149m-russian-sentence-70f5d3.md","json":"https://newruntime.com/posts/deepvk-released-user2-base-a-149m-russian-sentence-70f5d3.json"}
---

# DeepVK released USER2-base , a 149M Russian sentence encoder with 8,192-token context.

## Retrieval answer

DeepVK has released USER2-base, a 149-million-parameter Russian sentence encoder with an 8,192-token context window.

## Summary

DeepVK has released USER2-base, a 149-million-parameter Russian sentence encoder with an 8,192-token context window. Its Apache-2.0 weights, task-specific retrieval interface, and truncatable embeddings give teams a concrete model to evaluate for Russian-language search, RAG retrieval, clustering, and classification.

## What happened

DeepVK published USER2-base on Hugging Face as a new generation of its Universal Sentence Encoder for Russian. It is a 149M-parameter encoder built on RuModernBERT-base and fine-tuned for retrieval and semantic tasks. The model accepts inputs of up to 8,192 tokens and produces 768-dimensional representations. The repository is licensed under Apache-2.0 and lists Sentence Transformers, ONNX, Safetensors, Transformers, and text-embeddings-inference among its tags.

The release is aimed at sentence similarity and adjacent embedding tasks rather than text generation. The model card positions it for retrieval, reranking, classification, clustering, semantic textual similarity, and multilabel classification. It instructs users to add a task prefix to each input. `classification:` is presented as the general default; `clustering:` is recommended for grouping texts, discovering topics, or removing semantic duplicates. Retrieval uses separate `search_query:` and `search_document:` prefixes. DeepVK notes that prefix selection can vary by domain and that `search_query` may work better for some short-text classification tasks, while `search_document` can help long-context similarity work.

USER2-base also supports Matryoshka Representation Learning. Rather than requiring a separately trained small embedding model, users can truncate output vectors to one of the dimensions used in training: 32, 64, 128, 256, 384, 512, or 768. The provided Sentence Transformers example initializes the model with `truncate_dim=128`; the Transformers example takes the first dimensions after mean pooling and then normalizes the vector. This gives an implementer a direct way to test index size and retrieval-quality trade-offs.

DeepVK reports results on MTEB-rus and on the Russian subset of MultiLongDocRetrieval. On MTEB-rus, USER2-base has a reported mean task score of 61.12 and a mean task-type score of 59.59. Its category scores include 66.90 for retrieval and 74.28 for semantic textual similarity. In the model card's long-context comparison, restricted to models with an 8,192-token context length, USER2-base reports nDCG@10 of 54.17 on MLDR-rus. The card compares that result with USER2-small at 51.69, USER-bge-m3 at 58.53, KaLM-v1.5 at 53.75, jina-embeddings-v3 at 49.67, and E5-mistral-7b at 52.40.

The documented training path has three stages: RetroMAE, weakly supervised fine-tuning, and supervised fine-tuning. DeepVK says it used English and Russian data for cross-lingual transfer, including in-house parallel corpora, and mined 50 million non-overlapping text-block pairs from its cultura_ru_edu corpus. These are developer-supplied training details, not an independently reproducible evaluation of the resulting model.

## What the archive adds

We have not covered this before; it opens a new thread.

## Trend signal

No supplied Pattern or archive assumption genuinely fits this release. The evidence describes an embedding model and its evaluation rather than a change in agent architecture, interoperability, security, or organizational practice.

## Why it matters

For a Russian-language AI product, the embedding model often determines whether retrieval finds the right evidence before a generator sees it. USER2-base is notable less because it makes a universal quality claim than because its operating choices are explicit: a 149M model, an 8,192-token input limit, prescribed query and document prefixes, and seven supported vector sizes. That makes it possible to run a meaningful local evaluation instead of guessing how the model should be called.

The long context can reduce the need to fragment documents aggressively before embedding, while Matryoshka truncation can shrink storage and vector-search cost. Neither benefit is automatic. A team needs to assess whether longer passages improve retrieval on its corpus, whether the chosen prefix matches its task, and how much ranking quality it loses at a smaller dimension. The release provides an accessible starting point for those experiments.

## What happens next

- **Teams serving Russian-language retrieval will test USER2-base against their current embedder.** (3 months, 65%). Watch for: public benchmark reports or integrations that specify MTEB-rus, MLDR-rus, or private-corpus results.
- **Some deployments will use truncated vectors rather than the full 768 dimensions.** (6 months, 60%). Watch for: published index-size, latency, and quality comparisons at 128, 256, or 384 dimensions.

## Limits and uncertainty

The available evidence is a model card from the publisher. Its benchmark scores, training claims, and recommendations have not been independently verified here. MTEB-rus and MLDR-rus do not establish performance on a team's documents, queries, dialects, legal terminology, or production latency requirements. The model card reports comparison scores but does not establish identical serving conditions, hardware cost, throughput, or operational reliability across all listed models. An 8,192-token maximum is also an input capacity, not proof that embedding full-length documents yields better search than chunked indexing. The material does not provide a release date, adoption figures, or a hosted-service price.

## Skill to build now

Build a small Russian retrieval evaluation this week. Select 50 to 100 real user queries and label the documents that should be retrieved. Embed queries with `search_query:` and corpus passages with `search_document:`, then compare the top results with your current model. Repeat at 768, 384, and 128 dimensions, recording recall, nDCG, index size, and latency. Add a second test that compares your existing chunk length with longer passages, since the 8,192-token limit changes what you can reasonably index. Keep prefix and normalization choices fixed within each comparison.

## Reader takeaway

USER2-base is an open Russian embedding model that is straightforward to test: it supports long inputs, explicit retrieval prefixes, and several vector sizes from one set of weights. Treat the published benchmark as a hypothesis, then validate prefix choice, dimensionality, and document segmentation against your own retrieval workload.

## Open question

How much of the reported long-context retrieval performance survives on real Russian enterprise corpora once index cost, latency, and document-chunking policy are held constant?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)