We're proud to introduce Bleta-Logjike, an important development for Albanian capable large language models. Fine-tuned by the Dialogo team, Bleta harnesses the power of Gemma-3 architecture and transforms it through specialized training on multi-step reasoning and math datasets. This landmark model excels particularly in logical reasoning and step-by-step thinking, making it invaluable for complex problem-solving in the Albanian language.

Building Upon Previous Success

Bleta represents the evolution of our ongoing efforts to improve open-source Albanian language models. Following our successful release of Bleta-DeepSeek-R1 last year—our first Albanian model with enhanced logical reasoning capabilities—this new multilingual 27B parameter model marks a significant leap forward.

The new Bleta model maintains the Albanian language capabilities of its predecessor while substantially expanding its logical reasoning strengths. This allows it to perform complex reasoning tasks in Albanian with remarkable clarity and precision.

Revolutionary Logical Reasoning Training

What makes Bleta truly special is our efficient training approach. Drawing inspiration from the methodology used in DeepSeek's renowned logical reasoning models, we employed GRPO (Generative Reasoning-guided Policy Optimization) to fine-tune Bleta's thinking processes.

This sophisticated training process involved:

• Starting with a Strong Foundation: We built upon Google DeepMind's impressive Gemma 3 27B, which already exhibits excellent Albanian language capabilities and strong overall benchmark performance
• Step-by-Step Reasoning: Training the model to break down complex problems into logical steps before reaching conclusions
• Format Control: Teaching the model to clearly demarcate its reasoning process from its final answers
• Reward-Based Learning: Using sophisticated reward mechanisms to reinforce accurate logical deductions and problem-solving approaches

Bleta in Action: Logical Reasoning Examples

Step-by-Step Reasoning Logical Analysis Problem Solving Reasoning Chains

Step-by-step Thinking Process

Bleta breaks down complex problems into logical steps before reaching conclusions, making the reasoning process transparent and understandable. This example demonstrates how the model resolves multi-step logical challenges.

Logical Analysis Capabilities

Demonstration of how Bleta analyzes complex logical relationships and draws conclusions based on given premises. The model excels at identifying patterns and applying logical rules consistently.

Complex Problem Solving

Solving complex reasoning problems with precision is one of Bleta's core strengths. This example shows how the model tackles a challenging problem by methodically working through each step of the reasoning process.

Sophisticated Reasoning Chains

Bleta can build and follow sophisticated chains of logical deduction to arrive at accurate conclusions. This example demonstrates how the model connects multiple logical steps to solve a complex reasoning task.

Technical Specifications

# Using llama.cpp with GGUF format
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make
./main -m path/to/bleta-logjike-27b-finetune.Q8_0.gguf -n 1024 -p "Cila është zgjidhja e problemit të mëposhtëm? Nëse x + y = 10 dhe x - y = 4, sa është vlera e x?"

Supporting Albania's EU Integration Journey

Beyond its exceptional logical reasoning capabilities, future models in the Bleta line could prove invaluable in supporting Albania's EU integration process. The current model is an 8-bit quantized version of the 27B parameter model. With more Albanian language datasets, higher computing power, and improved training algorithms, we envision future Bleta releases to process and analyze complex texts more effectively.

The Bleta line of fine-tuned Albanian-capable LLMs can assist in the challenging process of transposing the EU acquis—the massive body of EU legislation comprising over 100,000 pages—into Albanian legal frameworks. Its logical reasoning skills help identify inconsistencies, compare regulatory approaches, and suggest potential harmonization strategies.

Real-World Applications

A Vision for Albania's AI Future

Bleta represents more than just a technological advancement—it's our latest effort to synthesize substantial recent developments in open-source LLMs for the benefit of Albanian speakers, businesses, and government. By enriching a benchmark-leading, open-source model with sophisticated logical reasoning capabilities, we're opening new possibilities across education, research, government, and business sectors.

Our roadmap includes further enhancing the Bleta line of fine-tuned LLMs in reasoning abilities, Albanian language skills, and real-world usefulness. As we rapidly move into the age of AI agents, we hope that our efforts will enable Albanian speakers, students, businesses, and government institutions to fully benefit from productivity gains without being tied to proprietary providers.

This groundbreaking project was developed by the Dialogo team as part of ongoing efforts to advance Albanian language technology. We welcome collaboration with researchers, educators, legal professionals, and government institutions interested in leveraging Bleta's powerful logical reasoning capabilities.