This blog publication focuses on new features and improvements. For a comprehensive list, including error repairs, please see Issue notes.
A new method based on Bethon to download models and inference
We have renewed how to download and use models to infer with a new Python -based method that gives priority to simplicity, speed and developer experience.
This flexible approach is designed with the user-focused user design, simplifying the work with models. It allows users to focus more on construction and repetition, and less on moving in API mechanics. The new method is to simplify the inference, accelerate the development, and greatly improves the ability to use in general.
Download the form
Clarifai Python SDK now makes it easy to download custom models. Whether you are using a pre -trained model of Face or Openai, or a layer of zero developed, the integration is smooth. Once downloaded, your model can immediately benefit from the powerful Clarifai platform features.
After importing, your model is published automatically and ready to use. You can evaluate it, or connect it to other models and agents operators in the workflow, or submit applications directly.
As part of this version, we greatly simplified how to determine model.py File to download the allocated form. New ModelClass The pattern allows you to implement predictand generateAnd streaming Methods without the need to strip or additional kettle. You can start a few lines of code.
Here is a quick example: a simple way that puts “hello world” for any introduction text, with integrated support for different types of broadcast responses. Check the full documents here.
Inference
The new reasoning approach provides an effective, developable and simple way to operate predictions with your models.
It is designed with the first focus of the snake, the developer, and reduces the complexity so that you can spend more time in construction and repetition, and a lesser time in dealing with the details of the low -level application programming interface.
Below is an example of how the customer side is made predict Summonary corresponds to predict The method specified in the previous section. Exit the documents here.
New published models
- Published Llama-4-SCOUT-17B-16E-InstructA strong model in the Llama 4 series includes 17 billion teachers and 16 experts to control advanced instructions. It supports a local context window of up to 10 million (8K supported on Clarifai), making it ideal for multi -accessible analysis, understanding of a complex code database, and smart work.
- Published QWEN3-30B-A3B-GGUFThe latest addition to the QWEN series. This new version is characterized by both dense and stitched models of experts (MEE), with significant improvements in thinking, following instructions, agent -based tasks, and multi -language capabilities. QWEN3-30B-A3B is outperforming larger models such as QWQ-32B, which benefits from fewer active parameters while maintaining strong performance through coding and thinking standards.
- The latest Openai has been published O3 Form, the strong and well -teacher LLM sets a new standard for performance through mathematics, science, coding and visual thinking tasks. It is designed for complex and multi -steering thinking and excels in solving technical problems, interpreting visual data such as plans and plans, high -risk decision -making, and creative thinking.
- Published O4-MiniA smaller model that has been improved for rapid and cost -effective thinking. Despite its built-in size, O4-MINI provides great accuracy on mathematics and coding standards such as AIME 2025. It is ideal for use situations that require strong thinking capabilities while maintaining cumin and low cost. Both models are also available on the field, try them here.
The stadium is strengthened expertise
- The discovery of the automatic situation is added based on the specified model – it now turns intelligently between the chat modes and the vision of forecasts.
- Improving the search for the model and identifying it for a faster and more accurate selection selection experience.
- The drop -down personal access (PAT) has entered, allowing users to easily enter their PAT keys into code scraps.
- The dynamic pricing display that is updated based on the specified publication has been implemented.
- The defined publishing identifier is now injected automatically into the inference code.
The control center has strengthened
Improving the community platform
- The exploration page is renewed with refreshing visual designs, display of distinctive models, and classified cases such as LLMS and VLMS.
- The individual model mode page has been updated with an improved user interface, direct access to the stadium, publishing lists, and additional improvements.
Additional changes
- The home page can now be accessed for all users, as it requires departments that require login automatically for unrestricted users. The new “modern activity” section of users shows their latest behavior and operations. We have also made improvements to use, performance and user experience in general.
- The accounts of the new institution now begin in the community plan by default, instead of inheriting the user’s personal plan. This change applies to users to society and basic and professional plans. The Users of the Foundation were not affected. The “member column” is displayed from now when a member joined the institution, and the settings pages are hidden from users without the required permissions.
- The bills section has been redesigned to try the most easy credit card management. We have added health verification to prevent refined card entries and support for setting up or changing the virtual credit card.
- Python SDK now supports the Bathonic models for an original experience. Fix the failure tests to improve stability. CLI is now 20x for most processes, including training contexts, improved error messages, and corrected return media in the creator. Learn more here.
Are you ready to start building?
With this first version of the Python, download and operation are inference on the customized models faster, simpler and easier than ever. Whether you combine a pre -trained or published model from zero point, Clarifai Python SDK gives you tools to move from the initial model to production with the lowest public expenditures.
explore Documents And start building today.
