CloudGoat/Mephisto-1.5-4B

Mephisto-1.5-4B: A Multi-Stage Merged Agentic Model

Mephisto-1.5-4B is a 4.5 billion parameter language model developed by CloudGoat, specifically engineered for agentic tasks. It leverages the Qwen3.5-4B architecture and a multi-stage merging methodology using mergekit to combine specialized capabilities.

Key Capabilities & Features

• Integrated Agentic Performance: Fuses reasoning, coding, and agentic functionalities into a single 4B checkpoint.
• Multi-Stage Merging: Utilizes a sophisticated three-stage pipeline:
  • Stage 1 (NuSLERP): Fuses reasoning-distilled and code-specialized models (e.g., Jackrong/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2 and Jackrong/Qwopus3.5-4B-Coder).
  • Stage 2 (DARE-TIES): Integrates agentic models for planning and general reasoning (e.g., BAAI/AREX-Turbo and InternScience/Agents-A1-4B), alongside a Japanese-focused model (Jackrong/Qwopus3.5-4B-v3).
  • Stage 3 (FrankenMerge/Passthrough): Structurally allocates layers for functional specialization, assigning specific layer ranges to base Qwen3.5-4B (for shallow features and Japanese), the agent foundation (for planning/function calling), and the reasoning/code fusion (for deep logic).
• Japanese Language Support: Incorporates models specifically for Japanese language modeling and general capabilities, targeting improved performance on Japanese benchmarks like ELYZA-tasks-100 and MT-Bench-JP.
• Consumer Hardware Deployment: Optimized for deployment on GPUs with 12GB VRAM, such as an RTX 3060.
• Apache-2.0 Licensed: All source models are Apache-2.0 licensed, allowing for commercial use of the merged model.

Good For

• Applications requiring a compact, agentic model with strong reasoning and coding abilities.
• Use cases demanding enhanced performance in Japanese language understanding and generation.
• Developers seeking a model deployable on common consumer-grade hardware.
• Research into multi-stage model merging techniques and their impact on specialized capabilities.