CloudGoat/Mephisto-4B-v2
CloudGoat's Mephisto-4B-v2 is a 4.5 billion parameter agentic language model based on Qwen3.5-4B, featuring a 32768 token context length. This model was developed using a multi-stage merging process with Mergekit, combining several fine-tuned models including Jackrong/Qwopus3.5-4B-v3, Jackrong/Qwopus3.5-4B-Coder, BAAI/AREX-Turbo, and InternScience/Agents-A1-4B. It is specifically designed for agentic applications, leveraging its complex merged architecture to enhance performance in such tasks.
Loading preview...
Mephisto-4B-v2: A Multi-Stage Merged Agentic Model
Mephisto-4B-v2 is a 4.5 billion parameter agentic language model developed by CloudGoat, built upon the Qwen3.5-4B architecture. It distinguishes itself through a sophisticated, multi-stage merging process using a forked version of Mergekit tailored for Qwen3.5. This approach aimed to create a more robust and capable agentic model compared to its predecessor.
Key Merging Details
The model's creation involved three primary merging steps:
- Step 1: Jackrong/Qwopus3.5-4B-v3 and Jackrong/Qwopus3.5-4B-Coder were merged using the SLERP method to form an intermediate model (Model 1).
- Step 2: BAAI/AREX-Turbo and InternScience/Agents-A1-4B were merged, also via the SLERP method, to create another intermediate model (Model 2).
- Step 3: The final Mephisto-4B-v2 was produced by merging Model 1 and Model 2 using the DARE-TIES method, integrating their respective strengths.
Intended Use
Mephisto-4B-v2 is specifically designed for agentic applications, where its complex merged architecture is expected to provide enhanced capabilities. Its 32768 token context length supports handling extensive interactions and complex problem-solving within agent frameworks.