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

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 11, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

The texdata/Qwen3.6-35B-A3B-Slovenian model, developed by MediaAtlas, is a 35.1 billion parameter Mixture-of-Experts (MoE) language model with 3 billion active parameters, fine-tuned for Slovenian language fluency, knowledge, and English-Slovenian translation. Based on Qwen/Qwen3.6-35B-A3B, it significantly improves Slovenian-LLM-Eval accuracy by +3.1 points and EN↔SL BLEU scores by +2.47 to +4.10 points compared to its base. This model also retains the vision capabilities of the base Qwen3.6, allowing it to process images, and is optimized for reasoning tasks.

Loading preview...

texdata/Qwen3.6-35B-A3B-Slovenian: Slovenian-Optimized MoE Language Model

This model is a 35.1 billion parameter Mixture-of-Experts (MoE) language model, with 3 billion active parameters, developed by MediaAtlas. It is a merged full model based on Qwen/Qwen3.6-35B-A3B, specifically fine-tuned for the Slovenian language through continued pre-training (CPT) and supervised fine-tuning (SFT).

Key Capabilities & Enhancements

  • Slovenian Language Proficiency: Significantly improves Slovenian fluency and knowledge, with an average accuracy gain of +3.1 points on the Slovenian-LLM-Eval benchmark (from 0.623 to 0.654).
  • Enhanced Translation: Achieves notable improvements in English ↔ Slovenian translation, with BLEU scores increasing by +2.47 for EN→SL (to 26.26) and +4.10 for SL→EN (to 34.95).
  • Reasoning Model: Designed to support reasoning tasks, with an enable_thinking option for direct answers or translation.
  • Multimodal (Vision): Retains the vision encoder from the base Qwen3.6 model, allowing it to process images, although its Slovenian training was text-only.
  • Efficient Loading: Optimized for 4-bit (nf4) quantization for efficient deployment, as the native bf16 forward for this architecture is noted as broken in current transformers.

Ideal Use Cases

  • Slovenian-centric applications: Chatbots, content generation, and knowledge retrieval in Slovenian.
  • English ↔ Slovenian Translation: High-quality translation services between the two languages.
  • Reasoning tasks: Applications requiring logical inference and problem-solving.
  • Multimodal applications: Projects that can leverage its image understanding capabilities alongside Slovenian text processing.

Finetunes
3 models