ModelMerge-JP/Mephisto-1.5-4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 10, 2026Architecture:Transformer0.0K Featherless Exclusive Cold

Mephisto-1.5-4B is a 4.5 billion parameter agentic language model developed by ModelMerge-JP, built upon the Qwen3.5-4B architecture with a 32768 token context length. It integrates reasoning, coding, and agentic capabilities through a multi-stage merging process using mergekit, combining specialized models for each domain. This model is designed to exceed the performance of Mephisto-4B-v2 on Japanese and agentic benchmarks while remaining deployable on consumer hardware like an RTX 3060 12GB.

Loading preview...

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

Mephisto-1.5-4B is a 4.5 billion parameter language model built on the Qwen3.5-4B architecture, specifically engineered for agentic tasks, reasoning, and coding. It leverages a sophisticated three-stage merging pipeline using mergekit to combine the strengths of several specialized models, aiming for enhanced performance in Japanese language understanding and agentic capabilities.

Key Capabilities

  • Integrated Agentic Intelligence: Fuses models specialized in planning, function calling, and general agentic reasoning.
  • Strong Reasoning & Coding: Incorporates a reasoning-distilled model and a code-specialized model to handle complex logic and code generation tasks.
  • Japanese Language Proficiency: Includes components optimized for general and Japanese language understanding.
  • Consumer Hardware Friendly: Designed to run efficiently on GPUs with 12GB VRAM, such as an RTX 3060.

Merging Methodology Highlights

  1. Reasoning-Code Fusion (NuSLERP): Combines a reasoning-distilled model with a code-specialized model from the Qwen3.5-4B family.
  2. Agent Foundation (DARE-TIES): Integrates three agentic models (BAAI/AREX-Turbo, InternScience/Agents-A1-4B, Jackrong/Qwopus3.5-4B-v3) for robust agentic behavior.
  3. Final Integration (FrankenMerge/Passthrough): Structurally allocates layers from the base Qwen3.5-4B, the agent foundation, and the reasoning/code fusion stages based on functional specialization.

Good for

  • Developing AI agents requiring robust planning and function calling.
  • Applications needing strong reasoning and code generation capabilities.
  • Projects targeting Japanese language understanding and generation.
  • Deployment on consumer-grade GPUs with limited VRAM (e.g., 12GB).

Finetunes
1 models