Flashtond22/Taltos-27B
Flashtond22/Taltos-27B is a 27.8 billion parameter language model, fine-tuned from Qwen/Qwen3.8-27B, specifically optimized for natural and idiomatic Hungarian language generation and reasoning. It features a hybrid Gated DeltaNet + Gated Attention architecture and supports a large 262,144 token context window, along with multimodal image and video input capabilities. This model excels at producing high-quality Hungarian text, including complex reasoning and nuanced phrasing, making it ideal for applications requiring native-level Hungarian linguistic proficiency.
Loading preview...
Táltos-27B: Hungarian-Optimized Multimodal LLM
Táltos-27B is a 27.8 billion parameter language model developed by Flashtond22, built upon the Qwen/Qwen3.8-27B base model. It is specifically fine-tuned for native-quality Hungarian language generation, focusing on natural phrasing, Hungarian-specific reasoning, and the ability to express uncertainty. The model retains the base model's extensive world knowledge, mathematical and programming capabilities, and multimodal vision features.
Key Capabilities
- Exceptional Hungarian Language Proficiency: Produces text with correct Hungarian grammar, idiomatic expressions, and nuanced phrasing, avoiding literal translations.
- Advanced Reasoning in Hungarian: Capable of step-by-step reasoning and argumentation in Hungarian, with significant gains observed in secondary-school exam problems and correcting false premises.
- Large Context Window: Supports a substantial 262,144 token context, enabling processing of very long documents.
- Multimodal Input: Interprets both image and video inputs, leveraging the base model's vision capabilities.
- Thinking Mode: Features a toggleable
enable_thinkingmode for enhanced reasoning processes.
Performance and Training
In head-to-head comparisons on 55 Hungarian tasks, Táltos-27B was preferred in 55.5% of cases over its base model, as judged by an independent model (Gemma-4-31B). The fine-tuning process involved a proprietary three-stage pipeline that separates content quality from Hungarian phrasing, allowing for maximal optimization of both. The entire training was conducted on a single NVIDIA L40S GPU.
Use Cases
- High-Quality Hungarian Content Generation: Ideal for creating letters, summaries, transcripts, and official documents in Hungarian.
- Complex Hungarian Reasoning Tasks: Suitable for applications requiring detailed logical deductions and argumentation in Hungarian.
- Multimodal Hungarian Applications: Can interpret visual information and generate Hungarian descriptions or analyses.
Limitations
While excelling in Hungarian, the model's world knowledge, math, and coding abilities are inherited from the base model and were not specifically enhanced. It lacks internet access, so information on recent events may be outdated. Users should verify specific numbers, dates, and rare names, and avoid using it as the sole basis for critical decisions (medical, legal, financial).