SpeakLeash.org announced today that their latest Polish AI chatbot model, Bielik v2, is now available to everyone. This long-awaited launch is a significant step forward in the development of AI in Poland, offering users advanced capabilities for interacting in Polish. Bielik v2, a collaboration between the SpeakLeash Foundation and the Academic Computer Centre CYFRONET AGH, stands out not only for its impressive 11 billion parameters , but also for its wide context window that allows for the processing of longer and more complex texts. Thanks to this model, the Polish community gains a tool that can significantly impact the development of local AI applications, contributing to the decentralization of technology and increasing its accessibility for all.

Contents

Key Features of the Bielik v2

It’s worth starting with the fact that Bielik, as the name might suggest, is a fully Polish product (more on that in a moment). Bielik v2 is a modern AI model that introduces a number of innovations aimed at improving the quality of interaction in Polish. Below are the key features that distinguish this model from others:

  1. Increased size : The model has 11 billion parameters, which significantly improves its ability to understand and generate text in Polish. The larger number of parameters allows for more complex linguistic analyses and better context capture.
  2. Wide context window : Bielik v2 supports context up to 32,768 tokens, which allows processing longer texts and more complex tasks. This functionality is especially useful in applications that require analysis of larger text fragments, such as summarizing articles or analyzing documents.
  3. Improved training data : The model has been trained on a much larger amount of high-quality data, which translates into better results on a variety of language tasks. High-quality training data is crucial to achieve accuracy and precision in the generated answers.
  4. Improved NLP : Bielik v2 performs better on Natural Language Processing (NLP) tasks such as text summarization, Naming Entity Recognition (NER), and question answering. These improvements make the model more versatile and effective for a variety of applications.
  5. Flexibility of implementation : The model is available in quantized versions, which allows its effective use on various hardware platforms. This allows users to adapt the model to their needs, regardless of the available computational resources.
  6. High quality of generated responses : Despite its relatively small size, Bielik v2 matches the quality of larger models, as confirmed by benchmark results such as Open LLM PL Leaderboard. The model is competitive with much larger systems, making it an attractive option for local applications.

What distinguishes Bielik v2 according to the chatbot itself?

I asked Bielik v2 what sets it apart from other AI chatbots, and its answer reveals a few key features that it highlights on its own. Here’s what it told me:

  1. Friendliness and Politeness : Bielik tries to conduct conversations in a friendly manner, avoiding inappropriate or offensive topics, in order to create a comfortable atmosphere for the user.
  2. Extensive knowledge : The model has extensive knowledge in various fields, from mathematics and programming to creative writing, making it a versatile interaction tool.
  3. Honesty : Bielik informs users when it does not know the answer to a question or when the information may be uncertain, demonstrating its transparency.
  4. Adaptation to the user's language : The model responds in the language the user uses, making communication easier.
  5. Accuracy and detail : For more complex questions, Bielik provides accurate and detailed answers, while for simpler questions he provides concise information.
  6. Safety : Bielik's priority is user safety, which is why it asks questions or informs about its concerns.
  7. Use of Markdown and Latex : The model uses Markdown to format text and Latex to represent mathematical formulas, which increases the clarity of the answers.
  8. Limitations : The White-tailed Eagle is aware of his limitations and is unable to learn on the fly, but he tries to be as helpful as possible within his abilities.

Let's emphasize this once again. The above features are listed by Bielik independently . They can build a positive image of its functionality. However, reality may be different , and users should see for themselves how the model works in practice. What are the real experiences of interacting with Bielik? This question remains open and encourages personal testing of its capabilities.

Running Bielik v2 locally or online

Bielik v2 offers users the flexibility to access their AI model, allowing them to run it both locally and online. For those with the right hardware, there is the option of using the Ollama tool , which allows you to run the model on your own computer. Simply type the command:

 ollama run SpeakLeash/bielik-11b-v2.2-instruct:Q8_0 <code readonly="true">ollama run SpeakLeash/bielik-11b-v2.2-instruct:Q8_0</code> ollama run SpeakLeash/bielik-11b-v2.2-instruct:Q8_0
			

This command starts Bielik v2, which gives the user the opportunity to test its functions in comfortable local conditions. However, it requires a computer with sufficient parameters for the model to run smoothly and efficiently.

For those who don't have powerful hardware or prefer faster access to the model, there is also an option to use the online platform. This can be done by visiting chat.bielik.ai . This option allows you to immediately test the capabilities of Bielik v2 without having to install additional software. Users can easily start interacting with the chatbot, making it accessible to a wider audience, regardless of their hardware resources.

Who are the creators of Bielik - SpeakLeash Foundation and ACK Cyfronet AGH

The premiere of Bielik v2, the latest AI-based chatbot, is the result of cooperation between the SpeakLeash Foundation and the Academic Computer Centre CYFRONET AGH. These organizations joined forces to create an advanced tool for interaction in Polish.

SpeakLeash Foundation (also called Spichlerz) is an open-source project that focuses on the development of Polish artificial intelligence. The foundation team carefully selected and processed Polish text corpora that were used to train the model .

In turn, ACK Cyfronet AGH is a high-performance computing center that has provided infrastructure for scaled processing in Polish. Support from the computing grant no. PLG/2024/016951 has enabled the use of state-of-the-art technologies and computing resources on the Athena and Helios supercomputers .

Thanks to this unique collaboration, the Bielik v2 model stands out for its unique ability to understand and process the Polish language. It generates accurate responses and performs a variety of language tasks with high precision .

The launch of Bielik v2 is an important step in the development of Polish artificial intelligence. The SpeakLeash Foundation and ACK Cyfronet AGH prove that by joining forces, advanced AI models adapted to local needs can be created.

speakleash panel
The dashboard allows you to track SpeakLeash's progress, providing insights into data capacity, industry discrepancies, and more.
📸: https://speakleash.org/dashboard/

CHAT ARENA PL - A platform for testing and developing Polish language models

CHAT ARENA PL is a unique platform created by the SpeakLeash Foundation, whose goal is to develop AI competences in Polish. It is a kind of "battlefield" where users can compare the skills of different language models in answering questions or prompts.

How does CHAT ARENA PL work?

The platform consists of several key elements:

  1. Typing prompts : Users start by typing a question or task for the AI model
  2. Generating responses : The system generates responses from two language models based on the given prompt
  3. Answer Evaluation : Users evaluate which answer is better. After evaluation, the system reveals the models used
  4. Prompt Saving : All entered prompts are saved for analytical purposes and to improve the quality of future models.

This approach not only engages users, but also allows them to directly compare the quality of answers. Many people may be surprised by the high quality that Bielik can provide, making it a valuable tool for learning and exploring AI in Polish.

CHAT ARENA PL Features

  • "Battle!" tab : The actual arena of language models, where users provide prompts and models generate answers
  • "Tasks" tab : Sample prompts that can serve as inspiration for users
  • "Leaderboard" tab : ELO ranking of models taking part in the confrontation
  • "Bielik vs. the world" tab : Possibility to compare the quality of texts generated by the Bielik.AI model with other models from around the world

Below is a screenshot from the arena. The prompt from the available suggestions is (original spelling):

 "If you had a dog, how often would you take it for a walk ?" 

Arena generated the following two answers. Think for yourselves which of the following answers is better. Below the image is information which answer was generated by Bielik.

chat arena
Click to see which models generated the responses

Answer A was generated by the gpt-4o-mini model
Answer B was generated by Bielik-2.1-11B

The goal of CHAT ARENA PL

The main goal of the platform is to develop AI competences in Polish. All entered prompts are used to analyze and improve the quality of future language models. The platform also serves to position models in relation to each other in the ELO ranking, which allows for a reliable comparison of their skills in tasks in Polish.