texdata/Qwen3.6-35B-A3B-Slovenian

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 11, 2026License:mixed-see-provenanceArchitecture:Transformer Featherless Exclusive Cold

The texdata/Qwen3.6-35B-A3B-Slovenian model is a 35 billion parameter Qwen3.6-A3B architecture, developed by texdata, specifically enhanced for Slovenian language fluency, knowledge, and English-Slovenian translation. This merged full model (bf16) was created by continued pre-training and supervised fine-tuning of the base Qwen/Qwen3.6-35B-A3B. It demonstrates improved performance on Slovenian-LLM-Eval benchmarks and significantly better en↔sl translation BLEU scores compared to its untuned base, making it ideal for Slovenian-centric NLP applications.

Loading preview...

texdata/Qwen3.6-35B-A3B-Slovenian: Slovenian-Enhanced Qwen3.6-35B-A3B

This model is a 35 billion parameter Qwen3.6-A3B architecture developed by texdata, specifically fine-tuned for the Slovenian language. It integrates a LoRA adapter, texdata/Qwen3.6-35B-A3B-Slovenian-LoRA, into the base Qwen/Qwen3.6-35B-A3B model through continued pre-training (CPT) and supervised fine-tuning (SFT).

Key Capabilities & Enhancements

  • Slovenian Fluency and Knowledge: Significantly improved understanding and generation in Slovenian.
  • Enhanced Translation: Achieves notable gains in English↔Slovenian translation, with BLEU scores increasing by +2.47 (en→sl) and +4.10 (sl→en) compared to the base model.
  • Reasoning Model: Designed to perform reasoning tasks, with an option to disable "thinking" for direct answers.
  • Benchmark Performance: Shows an average improvement of +3.1 points on the Slovenian-LLM-Eval acc_norm benchmark, winning 6 out of 7 tasks.
  • MTP Preservation: Multi-token-prediction (MTP) tensors are preserved for speculative decoding.

Important Usage Notes

Due to a known issue with the qwen3_5_moe architecture in current transformers versions, this model must be loaded in 4-bit (bitsandbytes nf4) or used via its GGUF build for correct functionality, despite its bf16 weights. The model is licensed under Apache-2.0 for its base, with CPT data (Slovenian Wikipedia, FineWeb2-sl) requiring attribution under CC BY-SA 4.0 and ODC-BY 1.0 respectively.

Finetunes
3 models