Dolphin 3.0

Dolphin 3.0 is the third generation of Eric Hartford's Dolphin series of uncensored instruct-tuned models, released in early 2025 on bases including Mistral Small 24B, Llama 3.1 8B, and Qwen 2.5 variants. The design philosophy is steerability: no imposed system prompt, no baked-in refusal behavior, and full control of alignment left to the person deploying it, which made Dolphin the reference local model for coding, roleplay, function calling, and agentic experiments.

An R1 reasoning variant trained on 800k reasoning traces extends the line into chain-of-thought use cases. The uncensored positioning carries obvious dual-use implications, and the base licenses (Llama 2 community license for Llama-based builds) constrain commercial redistribution. As a research artifact for alignment, controllability, and local inference, it remains one of the most downloaded open model families.