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 is a 4.5 billion parameter Slovenian assistant LLM, based on the Qwen3.5-4B architecture, designed for fully-local, on-device use. It excels in fluent Slovenian chat, bidirectional EN↔SL translation, and native tool/function calling with reasoning capabilities. The model is optimized for desktop deployment via GGUF (llama.cpp/LM Studio) and offers strong performance in Slovenian language tasks.

Loading preview...

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

Sokol is a compact 4.5 billion parameter language model developed by texdata, specifically engineered for fully-local, on-device deployment on desktops using GGUF (llama.cpp / LM Studio). Built upon the SicariusSicariiStuff/Qwen3.5-4B_Abliterated base, it has undergone extensive continued pretraining on ~1.78 billion tokens of Slovenian text for deep language adaptation, followed by supervised fine-tuning for diverse assistant capabilities.

Key Capabilities

  • Fluent Slovenian Chat: Engages in formal and natural conversations in Slovenian.
  • Bidirectional Translation: Capable of translating between English and Slovenian (EN↔SL).
  • Reasoning: Incorporates a <think> mechanism, distilled from a larger Qwen3.8-Max teacher model, allowing for internal reasoning traces before generating responses. This can be routed to reasoning_content to prevent it from being spoken in voice applications.
  • Native Tool/Function Calling: Supports Qwen3.6 XML format for native tool calls, enabling integration with external functions.
  • Identity: Presents itself as Sokol, a local Slovenian assistant.

Performance and Training

Sokol demonstrates significant improvements over its base model on the Slovenian-LLM-Eval benchmark, winning 6 out of 7 tasks and tying one, with an average accuracy increase of +0.047. Its continued pretraining halved Slovenian perplexity from 6.996 to 3.438 on held-out text. The model's training involved a mix of Slovenian Wikipedia, FineWeb2, and various SFT datasets for chat, translation, tool-calling, and reasoning distillation.

Good For

  • Developers building local, privacy-focused Slovenian AI applications.
  • Use cases requiring fluent Slovenian conversation and translation.
  • Integrating tool-use and reasoning into on-device assistants.
  • Applications where GGUF compatibility and efficient local execution are critical.

Finetunes
1 models