dphn/Dolphin3.0-Mistral-24B
TEXT GENERATIONConcurrency Cost:2Model Size:24BQuant:FP8Ctx Length:32kPublished:Feb 2, 2025Architecture:Transformer0.1K Cold

Dolphin3.0-Mistral-24B is an instruct-tuned model developed by Eric Hartford, Ben Gitter, BlouseJury, and Cognitive Computations, part of the Dolphin 3.0 series. This general-purpose model is designed for local deployment, offering capabilities across coding, mathematics, agentic tasks, and function calling. It emphasizes user control over system prompts and alignment, making it suitable for applications requiring custom behavior and data privacy.

Loading preview...

Dolphin 3.0 Mistral 24B Overview

Dolphin 3.0 Mistral 24B is a general-purpose instruct-tuned model developed by Eric Hartford, Ben Gitter, BlouseJury, and Cognitive Computations. It is part of the Dolphin 3.0 collection, aiming to provide a versatile local model for various applications.

Key Capabilities & Differentiators

  • General Purpose: Designed for a wide range of tasks including coding, math, agentic functions, and function calling.
  • User Control: Unlike proprietary models, Dolphin 3.0 allows users to fully control the system prompt and alignment, ensuring custom behavior and data privacy.
  • Steerability: Users can define the model's ethics and guidelines, making it adaptable to specific application requirements without external interference.
  • Low Temperature Recommendation: Experimental observations suggest optimal performance with a low temperature setting (0.05 to 0.1).
  • ChatML Format: Utilizes the ChatML format for chat templates, enabling clear instruction and system prompt definition.

Ideal Use Cases

  • Custom AI Applications: Businesses and developers needing full control over model behavior, alignment, and data handling.
  • Local Deployment: Suitable for scenarios where data privacy and offline operation are critical.
  • Specific Task Optimization: Can be tailored for specialized roles (e.g., a Golang coding assistant) through custom system prompts.

This model is built upon contributions from various open-source datasets and acknowledges the foundational work of Meta, Qwen, and OpenCoder.