About

Dolphin AI is an AI lab that develops uncensored open-source language models and operates Dolphin Network, a decentralized inference network that routes AI workloads to idle consumer and data center GPUs. Node operators contribute GPU capacity and earn POD tokens, which are also the payment unit for inference on the network. The project is built by Eric Hartford and the Cognitive Computations research collective, with the token deployed on Base.

History

Dolphin began in May 2023 as an open-source model project rather than a crypto network. Eric Hartford, reading Microsoft Research's Orca paper on progressive learning from GPT-4 explanation traces, concluded that the dataset behind such a model was unlikely to be released publicly and set out to reproduce it in the open. He published Dolphin as an uncensored, commercially licensed instruct dataset and an accompanying series of instruct-tuned models, paired with a thesis that alignment is a property the system owner should control rather than something a model provider imposes uniformly on every downstream application. The argument was that closed providers control the system prompt, the model versions, the alignment, and the query data, and that businesses building on top of those models inherit all four dependencies. Andreessen Horowitz provided an early grant that funded the compute for the first Dolphin models.

The model work grew into Cognitive Computations, an open-source research collective, and the Dolphin series was successively fine-tuned from Llama, Mistral, Qwen, and Phi base models. The models found commercial pull outside of crypto: Dolphin's uncensored models were adapted in collaboration with Venice.AI, whose consumer product serves them to end users.

The pivot into onchain infrastructure came from the supply side of AI inference. Dolphin Network is a peer-to-pool system rather than a rental marketplace: gamers and other consumer GPU owners contribute hardware that sits idle most of the time, the network routes inference requests to that pool, and operators earn tokens they can spend on inference or sell. Because inference demand is not latency-sensitive in the way most DePIN workloads are, supply and demand do not need to be geographically close, which distinguishes it from proximity-bound physical networks. The central engineering problem is preventing operators from modifying the node software to cheat, whether by returning pre-generated text instead of running inference or by silently substituting smaller or more heavily quantized models. The network addresses this with a combination of software integrity checks, randomly sampled validation of outputs, validator scoring, and cryptoeconomic bond deposits.

• May 2023: Dolphin is announced as an open-source, commercially licensed uncensored instruct dataset and model series based on the methodology in Microsoft's Orca paper
• December 2023: Dolphin 2.5 Mixtral 8x7B is released
• April 2024: Dolphin 2.9 is released, fine-tuned from Meta's Llama 3 8B
• January 2025: Dolphin 3.0 ships across Llama 3.1, Llama 3.2, and Qwen 2.5 base models
• February 2025: Dolphin 3.0 Mistral 24B and the R1 reasoning variant are released
• 2025: Dolphin Mistral 24B Venice Edition is co-released with Venice.AI
• October 2025: Dolphin Network beta goes live on Base, with GPU operators earning token rewards for serving inference
• March 2026: the token completes a 1:1 migration to a new Base contract and the ticker changes from DPHN to POD
• 2026: Dolphin Network V2 ships, rebuilt in Go with improved routing, load balancing, and auto-updating nodes