Tuguberk/Kizagan-E4B-Turkish-Agent-FunctionCalling-Hermes

VISIONConcurrency Cost:1Model Size:7.9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 1, 2026License:gemmaArchitecture:Transformer Cold

The Tuguberk/Kizagan-E4B-Turkish-Agent-FunctionCalling-Hermes model is a 7.9 billion parameter Gemma4-based language model, fine-tuned for Turkish function calling within agentic systems. It excels at multi-turn, conversational function calling scenarios, interpreting tool results, and generating Turkish responses. Optimized for agentic workflows, it translates Turkish natural language requests into structured tool calls for automation and chatbot applications.

Loading preview...

Model Overview

The Tuguberk/Kizagan-E4B-Turkish-Agent-FunctionCalling-Hermes is a 7.9 billion parameter model based on the Gemma4 architecture, specifically fine-tuned for Turkish function calling in agentic systems. It extends the capabilities of the AlicanKiraz0/Kizagan-E4B-Turkish-Reasoning-Model by adding robust tool-calling behavior, particularly for multi-turn, conversational scenarios.

Key Capabilities & Features

  • Turkish Function Calling: Converts Turkish natural language requests into structured tool calls (e.g., JSON). It maps Turkish user input to English API schema arguments.
  • Agentic Workflow Integration: Designed as a decision/action layer for agents, enabling Turkish assistants, chatbots, and automation workflows to interact with APIs.
  • Multi-Tool Support: Capable of generating multi-tool call chains for multi-step, dialogue-based tasks.
  • High Accuracy in Target Distribution: Achieves 96.4% function name accuracy and 94.9% argument key matching on unseen, Hermes-style Turkish function calling data, significantly outperforming the base model.
  • Preserved General Knowledge: Maintains the base model's general knowledge and comprehension (MMLU scores are comparable), with only a noted drop in free-form mathematical reasoning (GSM8K).

Use Cases & Strengths

  • Ideal for: Turkish multi-turn, conversational function calling in assistants and chatbots; agentic loops that interpret tool results and generate Turkish responses; typical API integrations (weather, calendar, search) for Turkish agents; multi-step, dialogue-based tasks involving tool chains.
  • Limitations: Less suitable for single-turn, highly technical, or out-of-distribution tool sets; scenarios expecting direct calls without dialogue (model may ask for clarification); free-form mathematical reasoning (a calculator tool is recommended instead); English function calling (untested).

Finetunes
2 models