texdata/Sokol-4B-SLO

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 31, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Sokol-4B-SLO by texdata (MediaAtlas) is a 4.5 billion parameter multimodal language model based on the Qwen3.5-4B architecture, fine-tuned for Slovenian language tasks. It excels at fluent Slovenian chat, bidirectional EN↔SL translation, reasoning, and native tool calling, designed for fully-local, on-device deployment via GGUF. The model also supports image input, leveraging its base model's vision encoder.

Loading preview...

Sokol-4B-SLO: On-Device Slovenian Assistant

Sokol-4B-SLO is a 4.5 billion parameter multimodal language model developed by MediaAtlas, specifically engineered for fully-local, on-device use on desktops via GGUF (llama.cpp / LM Studio). Built upon the SicariusSicariiStuff/Qwen3.5-4B_Abliterated base, it underwent extensive continued pretraining on approximately 1.78 billion Slovenian tokens, significantly reducing Slovenian perplexity by 51%. This adaptation makes it highly proficient in the Slovenian language.

Key Capabilities

  • Fluent Slovenian Chat: Engages in formal and natural Slovenian conversations.
  • Bidirectional Translation: Performs accurate English ↔ Slovenian translation.
  • Reasoning: Utilizes an internal <think> mechanism, distilled from a larger Qwen3.8-Max teacher model, for complex task processing.
  • Native Tool Calling: Supports Qwen3.6 XML format for function calls, enabling integration with external tools.
  • Multimodal (Vision): Capable of processing image inputs, with its vision encoder inherited from the base model.

Performance and Training

Evaluation on the Slovenian-LLM-Eval benchmark shows Sokol outperforming its base model across 6 out of 7 tasks, with a +0.047 average accuracy gain. Its training involved a multi-phase approach: continued pretraining on a large Slovenian text corpus (Wikipedia, FineWeb2) and supervised fine-tuning (LoRA) on diverse datasets covering chat, translation, tool-calling, and reasoning. The model presents itself as Sokol, a dedicated local Slovenian assistant.