TSjB/QM-4B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Jan 21, 2026License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Warm

TSjB/QM-4B is a 4 billion parameter language model based on Qwen3-4B-Instruct-2507, specifically fine-tuned to provide robust support for the Qarachay-Malqar language. It features an expanded tokenizer with increased Qarachay-Malqar token representation and a 40960-token context length. This model excels in generating text in Qarachay-Malqar, Russian, and English, making it ideal for applications requiring multilingual capabilities with a focus on the Qarachay-Malqar language.

Loading preview...

QM-4B: Qarachay-Malqar Language Support

QM-4B is a 4-billion parameter language model developed by TSjB, built upon the Qwen3-4B-Instruct-2507 architecture. Its primary differentiator is its specialized fine-tuning and tokenizer expansion to provide comprehensive support for the Qarachay-Malqar language (къарачай-малкъар тил), alongside strong capabilities in Russian and English.

Key Capabilities

  • Enhanced Qarachay-Malqar Support: Features an expanded tokenizer with significantly increased representation for Qarachay-Malqar symbols, improving linguistic accuracy and fluency.
  • Multilingual Generation: Capable of generating text in Qarachay-Malqar, Russian, and English, with support for other languages from the base Qwen3 model.
  • Extended Context Length: Offers a substantial context window of 40960 tokens, allowing for processing longer inputs and maintaining coherence over extended conversations or documents.
  • Optimized Training: Underwent a multi-stage training process including tokenizer expansion, embeddings-only training, and full fine-tuning of all model layers.

Good for

  • Qarachay-Malqar Language Applications: Ideal for chatbots, content generation, translation, and research focused on the Qarachay-Malqar language.
  • Multilingual Communication: Useful in scenarios requiring interaction across Qarachay-Malqar, Russian, and English.
  • Text Generation: Suited for tasks involving creative writing, summarization, and question-answering in its supported languages.

Limitations

  • The model was fine-tuned on text data (continued pretraining) rather than dialogues, which may affect its conversational abilities.
  • It may occasionally switch between languages within a single response.
  • Additional instruction tuning is recommended for improved instruction following.

Finetunes
1 models