CloudGoat/Mephisto-4B-v2.1

Mephisto-4B-v2.1: A Multi-Stage Merged Agentic LLM

Mephisto-4B-v2.1 is a 4.5 billion parameter language model developed by CloudGoat, designed for agentic workflows, reasoning, and code generation, with a focus on Japanese language performance. Built on the Qwen3.5-4B architecture, it leverages a sophisticated three-stage mergekit pipeline to combine the strengths of several specialized models, making it suitable for deployment on consumer-grade GPUs (e.g., RTX 3060 12GB).

Key Capabilities & Features

• Agentic Functionality: Integrates planning, function calling, and agentic reasoning through a dedicated merge stage combining models like BAAI/AREX-Turbo and InternScience/Agents-A1-4B.
• Reasoning & Code Generation: Fuses a reasoning-distilled model with a code-specialized model using NuSLERP, enhancing its ability to handle complex logic and generate code.
• Japanese Language Proficiency: Incorporates a general/Japanese model (Jackrong/Qwopus3.5-4B-v3) and utilizes Qwen3.5-4B's base layers for Japanese language modeling.
• Multi-Stage Merging: Employs a unique pipeline involving NuSLERP for reasoning-code fusion, DARE-TIES for agent foundation, and FrankenMerge (Passthrough) for layer-wise functional specialization.
• Hardware Friendly: Optimized for efficient inference on consumer hardware, with GGUF quantizations available (Q8_0 at ~4.5 GB, Q4_K_M at ~2.6 GB).

Ideal Use Cases

Mephisto-4B-v2.1 is well-suited for applications requiring:

• AI Agents: Developing autonomous agents that can plan, execute tasks, and utilize tools.
• Code Assistance: Generating and understanding code, particularly in environments where reasoning is critical.
• Japanese Language Applications: Tasks involving Japanese instruction following, conversation, and natural language understanding.
• Resource-Constrained Deployments: Running advanced LLM capabilities on consumer GPUs or edge devices.