Model Overview
RefalMachine/RuadaptQwen3-32B-Instruct is a specialized adaptation of the Qwen/Qwen3-32B model, meticulously engineered for superior performance in Russian language tasks. Developed by RefalMachine, this 32 billion parameter model underwent a significant transformation involving a tokenizer replacement and extensive continued pre-training on a Russian corpus. A key innovation is the application of Learned Embedding Propagation (LEP), which contributes to its enhanced capabilities.
Key Differentiators & Capabilities
- Optimized for Russian Language: The model features an extended tiktoken cl100k tokenizer, augmented with 48,000 Russian tokens, specifically designed for the Russian language.
- Accelerated Generation Speed: Thanks to the new tokenizer and specialized training, the model achieves up to a 100% increase in Russian text generation speed compared to the original Qwen3-32B, depending on context length.
- Hybrid Reasoner: Like its base model, RuadaptQwen3-32B-Instruct operates as a hybrid reasoner, capable of complex thought processes. Users can toggle this 'thinking mode' on or off using specific tokens or API parameters.
- Continued Pre-training: The model benefited from continued pre-training on a substantial Russian-language dataset, followed by instruction tuning and quality calibration.
Recommended Usage
For stable performance, it is advised to use low temperatures (0.0-0.3), a top_p in the range of 0.85 to 0.95, and a repetition_penalty of 1.05. The model's responses are a reflection of its training data and do not represent the authors' opinions. Users should exercise caution, especially given its foundation on a third-party pre-trained model for which the current authors are not responsible for the initial pre-training.