astom-M/matsuo-llm-advanced-phase-f4b

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Feb 22, 2026Architecture:Transformer Cold

The astom-M/matsuo-llm-advanced-phase-f4b is a 7.6 billion parameter language model, merged using the DARE TIES method with Qwen/Qwen2.5-7B-Instruct as its base. This model combines two pre-trained models, prioritizing 'phase_d' with a 70% weight and supplementing with 'phase_e2b' at 30%. It is designed to leverage the strengths of its constituent models, offering a balanced performance profile for general language tasks with a 32K context length.

Loading preview...

Model Overview

This model, astom-M/matsuo-llm-advanced-phase-f4b, is a 7.6 billion parameter language model created through a sophisticated merging process. It utilizes the DARE TIES merge method, known for its effectiveness in combining pre-trained models while preserving their capabilities.

Key Characteristics

  • Base Model: Built upon the robust foundation of Qwen/Qwen2.5-7B-Instruct, ensuring strong general language understanding and generation.
  • Merge Composition: It is a blend of two distinct pre-trained models: ./outputs/phase_d/merged_model (70% weight) and ./outputs/phase_e2b/merged_model (30% weight). This specific weighting aims for a more conservative integration, with phase_d being dominant.
  • Merge Method: Employs the DARE TIES technique, which is designed to efficiently merge models by pruning and re-scaling weights.
  • Configuration: The merge process included parameters such as normalize: true and int8_mask: true, and was performed using bfloat16 dtype.

Potential Use Cases

Given its merge-based architecture and foundation on Qwen2.5-7B-Instruct, this model is suitable for a variety of applications requiring a capable 7B-class LLM, including:

  • General text generation and completion.
  • Instruction following and conversational AI.
  • Summarization and question answering tasks.