Contact UsGet started

Yandex Foundation Models release notes

Written by
Updated at April 21, 2025

Release as of 31/03/2025

  1. The RC branch now features the YandexGPT Lite 5th generation model, which supports contexts of up to 32,000 tokens in both synchronous and asynchronous modes.
  2. Added support for OpenAI tools to work with text generation models.

Release as of 19/03/2025

Increased some limits for AI Assistant API: now you can add up to 10,000 documents with the total size of 5 million tokens to a search index. You can upload up to 100 documents at a time. Also, increased the maximum number of threads to 10,000 and maximum number of messages per thread to 100,000. For a complete list of limits, see Quotas and limits in Yandex Foundation Models.

Release as of 25/02/2025

The YandexGPT 5th generation model is available for testing (RC branch). The 5th generation key upgrades include:

  • Function calling was significantly improved.
  • Added support for structured output. This feature enables you to set up the model to generate responses in random JSON format or according to the provided schema. For more information on structuring model output, see Text generation overview.
  • Increased the supported context to 32,000 tokens for all modes.

Release as of 11/02/25

  1. Updated the Llama 70B model version. Now Llama 3.3. Llama was created by Meta. Meta is designated as an extremist organization and its activities are prohibited in Russia. is available in all branches.
  2. A new version is out: ML SDK 0.3.1. It features the following updates:
    • Python 3.8 is no longer supported.
    • Added the multipart mode for uploading large datasets.
    • Added the allow_data_logging option that allows you to use dataset data to improve the tuning service when loading datasets.
    • Added the validation_errors field that stores dataset validation errors.
    • Replaced the grpc_credentials field with verify in the SDK builder.

Release as of 7/02/25

Added support for the reasoning mode in the YandexGPT Pro model.

Release as of 9/12/24

Upon request, LoRA-based model and classifier tuning has been added in Preview.

YandexGPT-based classifiers are now publicly available.

Release as of 4/12/24

Llama 3.1 models are now available in Foundation Models. For model usage costs, see Yandex Foundation Models pricing policy.

Release as of 2/12/24

The YandexGPT 4th generation model is now available in the main branch (Latest). Version 3 will remain available in the Deprecated branch as per the models' lifecycle.

Release as of 21/11/24

The AI Assistant API functionality is now available to all Yandex Foundation Models users at the Preview stage.

Release as of 01/11/24

  1. Image generation with YandexART is now publicly available. Starting November 1, 2024, YandexART is billed according to the rules described on the Foundation Models pricing policy page.
  2. Increased the YandexART quotas for the number of generation requests per minute and full day (24 hours).
  3. Increased the YandexGPT quota for the number of concurrent generations. For information on the restrictions in place, refer to Quotas and limits in Yandex Foundation Models.
  4. Starting December 2, 2024, the YandexGPT model's test version (RC branch) will become the main version (Latest branch). The current version will remain available in the Deprecated branch as per the models' lifecycle.

Release as of 24/10/24

  1. The YandexGPT 4th generation model is available for testing (RC branch). Compared to the previous generation, the model's response speed has increased by an average of 2.5 times. The maximum context the model operates has also been increased. In asynchronous mode, 4th generation models can process up to 32,000 tokens. And now there is the YandexGPT Pro 32k model added to process large contexts in synchronous mode. For more information on model limitations, see Quotas and limits in Yandex Foundation Models.
  2. Increased the maximum number of tokens per response in the management console.

Release as of 10/10/24

Updated the YandexART model. The updated version has better understanding of requests, considers more details, and can generate text in Latin characters on the image.

Release as of 21/06/24

Starting June 24, 2024, the YandexGPT Lite model based on YandexGPT 3 is available by default (latest branch). The deprecated model based on YandexGPT 2 is available in the deprecated branch until July 1, 2024.

Release as of 07/06/24

Updated the YandexART model:

  • Compared to the previous version, the updated model understands text requests better and creates more realistic images.
  • Added the optional aspectRatio parameter for image aspect ratio.

Release as of 29/05/24

Added the text classification feature.

Release as of 28/05/24

The YandexGPT 3-based YandexGPT Lite RC model is now available in Release Candidate status. The model may replace the current YandexGPT Lite in the future.

Release as of 19/04/24

Added the ability to send asynchronous requests to YandexGPT models fine-tuned in DataSphere.

Release as of 09/04/24

  1. Added generation of images based on text description. The YandexART model works in asynchronous mode and is available in the management console in Foundation Models Playground and via the API.
  2. Added examples of requests to YandexART in the documentation.

Release as of 25/03/24

Added a new YandexGPT Pro model of the YandexGPT 3 family.

Release as of 26/01/24

Updated the YandexGPT and YandexGPT Lite models and improved their response performance.

Release as of 06/12/23

  1. Added a new YandexGPT model for asynchronous mode.
  2. Significantly revised the service's API.
  3. Unified model names and the way to access the models.

Release as of 02/08/23

  1. Increased the total number of prompt and response tokens.
  2. Added a new mode called Chat.
  3. Added a method for counting the number of tokens in a request.
Previous
Public materials
Next
FAQ