CloudGoat/Mephisto-4B-0725

Mephisto-4B-0725: A Tool-Calling Enhanced Merge Model

Mephisto-4B-0725 is a 4 billion parameter language model created by CloudGoat through a sophisticated merge of pre-trained models, primarily based on Qwen/Qwen3.5-4B. Its core innovation lies in endowing the strong conversational abilities of InternScience/Agents-A1-4B with enhanced tool-calling capabilities, a feature often localized in specific model components.

Key Capabilities

• Agentic Behavior: Designed to integrate tool-calling functionality, which was previously lacking in its base conversational model.
• Preserved Conversational EQ: Utilizes the DARE-TIES merge method with a high drop rate to carefully extract task-relevant parameters, aiming to maintain the original model's diffuse conversational intelligence.
• Merge Method: Leverages the DARE TIES method, combining InternScience/Agents-A1-4B and allenai/tmax-4b to achieve its specialized functionality.

Benchmarks

While primarily focused on tool-calling, the model's general language understanding performance is comparable to its base InternScience/Agents-A1-4B model across several Japanese language benchmarks (JCommonsenseQA, JNLI, JSQuAD, MARC-ja), with minor variations. For instance, it scores 84.67% on JCommonsenseQA and 77.00% on JNLI. On GAIA Level 1, both models achieved 12/53.

Good For

• Applications requiring agentic behavior and tool-calling capabilities.
• Use cases where a balance between conversational fluency and structured interaction is crucial.
• Developers looking for a 4B parameter model with specialized function-calling integration.