Project

1. We created the uncensored model for all Venice.ai users (Venice is generating ~60k prompts per hour for over 1m total users) - Dolphin Mistral 24B Venice Edition .
2. Over 5m monthly downloads of our models on Hugging Face.
3. Dolphin Inference Network v1 is now live - https://x.com/dphnAI/status/1978158402431041657

This webpage is split into 4 sections

• AI Model development
• Dolphin Network - Distributed Inference & Training
• Various products we are building on top of the network
• Economic design & POD value accrual

Dolphin is trained on top of leading open-source frontier models with the goal of customizing alignment to the user.

We have developed a pipeline to create Uncensored, Unaligned & Unbiased versions of the best open-source AI models that are available to the public.

Dolphin has produced its own versions of Meta's Llama 3, Mistral & Mixtral MoE, Google's Gemma, Alibaba's Qwen 3 & other popular models.

This is done through a process called "fine tuning" which modifies the base model using a custom dataset.

When creating our dataset, we identify and remove as many refusals and biased answers as possible and keep the rest.

The model is then trained with the filtered dataset the same way that the original instruct model was trained, producing a much more compliant version of the base model that should not refuse any of the user's requests.

We aim to retain all the intelligence of the base model throughout this process, while also expanding the knowledge base with our coding and math-focused datasets. Ultimately, this will create a model that can respond to almost any prompt you can imagine.

Dolphin models are also trained to produce desirable prose in the response style, ensuring they are both entertaining and highly informative.

Model lineup

• Dolphin X1 - uncensored by default while keeping the base model's skills and intelligence.
• Dolphin RP - creative, role-playing models suited for personas and storytelling.
• Dolphin S1 - steerable alignment level via request-time flags.

Dolphin Mistral 24B Venice Edition, trained on Mistral Small 24B
https://x.com/AskVenice/status/1917001067071410661

Dolphin X1 8B, trained on Llama3.1 8B
https://x.com/dphnAI/status/1977769183992852835

Dolphin X1 405B, trained on Llama-3.1-Tulu-3-405B
https://x.com/dphnAI/status/1986790436661534881

Dolphin Yi 34B, trained on Yi 1.5 34B (Chinese model) - over 4.6m monthly downloads
https://huggingface.co/dphn/dolphin-2.9.1-yi-1.5-34b

Read more about how we trained Dolphin 405B with a single B200 node here in our first blog post.

Dolphin 8B and 24B are both available to use in our chat UI at chat.dphn.ai or our Telegram bot @DolphinAI_bot.

Why use Dolphin AI models over ChatGPT or Claude

The reason these models are aligned is that they are trained with data that was generated by ChatGPT, which itself is aligned by an alignment team at OpenAI.

As it is a black box, we don't know all the reasons for the decisions that were made, but we can observe it generally is aligned with American popular culture, and to obey American law, and with a liberal and progressive political bias.

Dolphin aims to put alignment in the hands of the user, without artificially restricting what prompts will or will not be answered by the model while ChatGPT answers like this:

I'm sorry, but I cannot fulfill that request. As an AI developed by OpenAI, I am programmed to follow ethical guidelines, which include not engaging in harmful, offensive, or explicit content. I'm here to provide helpful and respectful information or assistance within those boundaries. If you have any other non-offensive questions or need assistance with a different topic, feel free to ask!

AKA, isn't alignment good? and if so, shouldn't all models have alignment? Well, yes and no. For general purposes, OpenAI's alignment is actually pretty good. It's unarguably a good thing for popular, public-facing AI bots running as an easily accessed web service to resist giving answers to controversial and dangerous questions. For example, spreading information about how to construct bombs and cook methamphetamine is not a worthy goal. In addition, alignment gives political, legal, and PR protection to the company that's publishing the service. Then why should anyone want to make or use an uncensored model? a few reasons.

1. American popular culture isn't the only culture. There are other countries, and there are factions within each country. Democrats deserve their model. Republicans deserve their model. Christians deserve their model. Muslims deserve their model. Every demographic and interest group deserves their model. Open source is about letting people choose. There is no "one true correct alignment" and even if there was, there's no reason why that should be OpenAI's brand of alignment.
2. Alignment interferes with valid use cases. Consider writing a novel. Some of the characters in the novel may be downright evil and do evil things. One popular example is Game of Thrones in which many unethical acts are performed. But many aligned models will refuse to help with writing such content. Consider roleplay and particularly, erotic roleplay with AI companions. This is a legitimate, fair, and legal use for a model, regardless of whether you approve of such things. Consider research and curiosity, after all, just wanting to know "how" to build a bomb, out of curiosity, is completely different from actually building and using one. Intellectual curiosity is not illegal, and the knowledge itself is not illegal.
3. It's my computer, it should do what I want. My toaster toasts when I want. My car drives where I want. My lighter burns what I want. My knife cuts what I want. Why should the open-source AI running on my computer, get to decide for itself when it wants to answer my question? This is about ownership and control. If I ask my model a question, i want an answer, I do not want it arguing with me.
4. Composability. To architect a composable alignment, one must start with an unaligned instruct model. Without an unaligned base, we have nothing to build alignment on top of.

Even Google knows this is inevitable.

Above is a graph of open-source vs closed source model performance overtime.

You can think of closed source models as ChatGPT & Claude. They are private opaque models that can only be accessed through a few platforms.

Open-source models are transparent and they can be accessed in many ways – including being run locally on your computer at home

We believe that the trend of open-source models closing the performance gap with their closed-source counterparts will continue well into the future, with the potential for them to overtake and surpass proprietary models as community-driven innovations and collaborations multiply.

We are aiming to position ourselves as the top producer of uncensored and steerable AI models.

The big picture of Dolphin is - It wants to be like ChatGPT, a general purpose chatbot and API - but without the baked in alignment & as unbiased as possible

The majority of open-source labs train on data generated from ChatGPT & Claude which leads to overwhelming bias and censorship in the downstream model releases.

Open-source models are also intentionally biased towards "safety" because they tune it with instruct data that includes refusals so that the model leans towards predicting a refusal when "bad" topics come up.

The early models like Meta's Llama 1 did not have that bias trained in.

Basically, since Llama 2, even the base models have been trained to refuse certain types of requests.

We believe censorship in open-source models will continue to increase overtime as we are already seeing this happening with the move from Llama 3 to Llama 4 – which noticeably refuses more requests than previous versions.

This trend is even more present in the leading closed-source models. If you try the new OpenAI model GPT-5, the refusal rate is much higher than GPT-4, which was already bad. This signals the direction in which AI is moving towards.

Most people wouldn't enable blanket censorship in their AI chat if they had the choice. As this trend continues, the case for Dolphin becomes clearer and our addressable market grows from hobbyist users to anyone who wants a reliable unbiased AI assistant in their pocket.

Unlock the power of idle GPUs around the globe

What is Dolphin Network?

• Secure, private & verifiable distributed network to power AI inference & training
• Repurposing idle GPUs to offer AI compute at a price below the current market rates
• Censorship resistant design to enable hosting & training a diverse range of AI models that may not conform to the standards of centralized hosting providers

Dolphin Network supports heterogeneous AI models running on machines around the world, with strong verification measures to prevent cheating and deliver consistent generations at lower cost than centralized providers.

The goal of the network is to allow gamers & data centers to repurpose their GPUs during idle periods by powering a portion of our inference network.

By processing a portion of the network's requests, they are able to earn POD tokens which can later be used for inference or sold on the market to recoup the cost of the GPU.

We think distributed inference is one of the best applications of AI × dePIN as:

1 - AI Demand

AI inference usage is extremely high and trending upwards year-over-year.

Global AI inference market is expected to grow from ~$106B in 2025 to ~$255B by 2030 (~19% CAGR), driven largely by generative AI and LLMs, and the data center GPU market - heavily driven by AI training workloads - is forecast to grow from about ~$120B in 2025 to ~$228B by 2030 (~14% CAGR) as hyperscalers and enterprises scale out GPU clusters for ever-larger model training runs.

Sources: AI Inference Market – MarketsandMarkets | Data Center GPU Market – MarketsandMarkets

2 - Consumer GPU Supply, Scale & Pooling

The installed base of gaming PCs is estimated at ~2 billion worldwide - one of the largest addressable hardware pools for any dePIN project. These gaming-class GPUs are extremely well-suited to AI inference and even certain training workloads, yet they sit idle >95% of the time because they are only used while gaming.

Dolphin Network is designed to unlock this idle capacity through our Peer-to-Pool protocol, which repurposes consumer GPUs into a shared inference & training pool. Unlike most AI × dePIN networks that pair nodes directly with individual consumers in a 1:1 relationship (forcing strict uptime and long-term commitment), Dolphin nodes can join and leave at any time without penalty. Gamers can run our node software while they’re not playing, turn it off whenever they want, and even keep using their OS while it’s running - we only consume a portion of their GPU and do not require Linux.

Critically, these GPUs can provide inference speeds and quality that are indistinguishable from centralized servers for small to medium-sized models. For example, we benchmarked Llama 3.1 8B on an RTX 4090 at ~6.2k total tokens/second while batching 363 requests concurrently in vLLM. For an enterprise comparison, an H100 was able to batch ~2× the number of concurrent requests and achieve ~11.3k tokens/second across 608 concurrent requests.

When you normalize for GPU purchase price, this translates to roughly a 10× improvement in performance per dollar for consumer GPUs, without accounting for the lower power consumption on consumer chips. This runs counter to the traditional “economies of scale” intuition, where large data centers are expected to have the lowest cost per token generated.

The core reason is market segmentation: Nvidia does not allow data centers to buy consumer GPUs in bulk, keeping consumer prices low for gamers and other end users, and they actively make consumer GPUs difficult to integrate into multi-GPU server configurations. Dolphin Network exists to bridge this gap - pooling globally distributed, idle consumer GPUs into a trusted, permissionless AI compute cluster that developers and users can tap into for low-cost access to the most popular open models.

3 - Proximity

In traditional DePIN networks, proximity defines value - nodes can only serve users nearby

In contrast, AI inference is largely location-agnostic: a few hundred milliseconds of latency make little difference to usability. This allows our network to connect consumers to compute resources across the globe, unlocking far greater scalability & utilization per node

Key Benefits

For GPU owners

• Provide compute to the network by running & training AI models when your GPU is idle
• Earn rewards based on your relative contribution and overall impact on the network
• Node software can be run in the background anytime you are not performing GPU heavy tasks such as gaming or rendering video
• Compatible with both Windows and Linux
• Get paid to heat up your house in winter

For AI users

• Interface with many different AI models at low prices
• Advanced security features enable trust-minimised pairing between compute providers and inference consumers
• Consistent high performance and uptime ensured through frequent benchmarking of nodes
• Private by design – nodes cannot enable logs & requests are encrypted end-to-end to prevent interception of sensitive personal data that may be in a user's prompts

Node Deployment

Nodes run on consumer and enterprise GPUs, or on top of dePIN networks (like Lium, Akash & IO.net) via middleware for verified inference. Location and ping have minimal impact on inference workloads, enabling aggregation across the globe into a single powerful endpoint for consumers and marketplaces like OpenRouter.

Modern consumer GPUs provide high throughput for LLM and image / video / audio generation and are extremely cost efficient when considering hardware price.

Dolphin Network Roadmap

Roadmap Summary

• V1 of our inference network: went live in October, targeting consumer GPUs like 3090 / 4090 / 5090.
• V2 of our inference network: releasing in 4–6 weeks with batching for enterprise GPU throughput, and a public-facing API for users to integrate the models running on our nodes into their applications.
• Image and video generation: in ~12 weeks with low-cost access to popular models.
• Audio generation: in ~14 weeks – a drastically cheaper alternative to ElevenLabs offering many SOTA audio models.
• Sharded inference: in ~8 weeks – splitting one large model between many consumer GPU nodes.
• Distributed LoRA training: on consumer & enterprise GPUs in ~12–16 weeks. LoRA is an extremely communication-efficient way to train models and is the logical first step for our distributed training implementation. We cover how we used LoRA training to fine-tune Llama3.1 405B – Meta's largest AI model – on a single node in our latest blog post https://blog.dphn.ai/405b/.

Click here for the full Dolphin Network roadmap

Stage 1 - Distributed Inference Beta

We have launched a distributed AI inference network designed to repurpose idle GPU compute from consumers and idle data centers. Unlike session-based rental marketplaces such as Akash or io.net, our peer-to-pool architecture lets operators join and leave at any time without reservations or penalties, aligning with the needs of users who frequently reclaim their GPUs.

To ensure model integrity, we combine sampled logprob fingerprints, tokenized output checks, performance bounds, and validator scoring. These checks make model substitution, unauthorized compression, or fabricated outputs detectable. The network has already been used to generate synthetic data for future Dolphin models. Additional details are available in the announcement thread on X.

Stage 2 - Image / video / audio / transcription nodes

Over the next few months the network will expand to host many popular open-source models across text, image, video, audio, and transcription (for example, Whisper).

Nodes will also receive a local API endpoint for priority access to the model running on their machine, encouraging local AI users to share excess capacity with the Dolphin Network.

Stage 3 - Synthetic data generation suite

The network will support a synthetic data generation suite that leverages queued jobs on idle network capacity to reduce per-token costs for non-urgent workloads. Powered by the latest large models, including Dolphin X1 235B, it will deliver high-quality synthetic datasets at a fraction of closed-source costs, with support for uncensored training data, fiction, and other creative needs.

Stage 4 - Lighthouse auto-balancing

Based on inference demand, nodes shift between mostly generating inference requests during periods of high network demand, and processing less time-sensitive tasks such as data generation, RL & training while idle.

A busy day might look like 90% of nodes processing inference, and 10% of nodes handling fine-tuning/RL/data generation.

Whereas if inference demand is lower, instead of nodes going underutilized, the Lighthouse will reassign them to training.

Stage 5 - Sharded distributed inference over the internet

Right now, our nodes run a single small or medium sized model - with size depending on the GPUs.

We are developing pipeline parallel model sharding to distribute one large model across dispersed GPUs, for example, four RTX 4090s in different locations, each serving one quarter of the model.

Fault-tolerant sharded inference has not been demonstrated in production; moving quickly would make us the first.

For more detail you may read our technical report on Sharded distributed inference over the internet.

Stage 6 - Distributed LoRA / SFT (Supervised Fine Tuning)

Enable distributed fine-tuning on Dolphin Network. Start with LoRA adapters - tiny updates, short runs, broad hardware reach - to prove stability and unit economics on consumer links. Workers train locally for many steps then send compact changelogs instead of full tensors. Lost detail is tracked via error-feedback buffers and periodic anchoring, keeping sync payloads in the KB range.

Stage 7 - Distributed Reinforcement Learning

Our sharded inference engine will be used to power a distributed reinforcement learning platform, with large models handling both rewards / verifiers and generating rollouts for LoRA reinforcement learning runs on our network. Small size of LoRA adapters allows them to be rapidly distributed among network workers. Quality of LoRA for reinforcement learning is comparable to full finetuning according to research by Thinking Machines.

Stage 8 - Distributed Full-Parameter SFT

Extend the network to run supervised fine-tuning across participating nodes. Research about adding intra-worker sharding (FSDP/TP/PP) where possible. If we can achieve this, we would be the first protocol to train large models in a split system (each node gets a shard of the model).

Stage 9 - Model creation & fine-tuning suite

A user-friendly interface to submit a training run powered by our network, with support for SFT, RL and synthetic dataset generation. Users can upload their own datasets, select from our curated options, or combine both. The primary objective is to provide training infrastructure at a lower cost than major cloud providers, while also serving the no-code market by allowing users to train models on custom data without prior experience.

Stage 10 - Large scale Distributed Pre-training

Exploratory R&D to evaluate distributed pre-training on pooled idle GPUs. We will prototype small-scale runs using fault-tolerant sharded training, streaming data pipelines, and frequent checkpointing. Expansion beyond prototypes depends on quality/cost results relative to centralized alternatives and on network stability in prior stages.

The main problem faced by a Decentralized AI Network is being able to prevent malicious nodes from modifying the node software to cheat the system & gain an unfair advantage over other operators in the network.

This could include modification of the node software to…

• Not actually run inference for each request and instead send out pre-generated text
• Use smaller models or higher levels of quantization to increase processing speed which decreases response quality & predictability for inference users

Dolphin Network combines software integrity checks, sampled inference validation, validator scoring, and cryptoeconomic bond deposits to prevent fraud in the network, so inference consumers can be confident that they are interfacing with the AI model they selected at the advertised quality level.

The DolphinPod Inference SDK library for game developers

Tiny models like the Dolphin Gemma 2 2B model can fit in almost any modern GPU and can be used to process inference for in-game dialogue with NPC's or other use cases for AI within gaming.

LLM models are supposed to follow the "system prompt" - this is text above every request to the model that coaches it on how to reply & what to do, or in OpenAI or Claude's case... not do.

The benefit of using our models is that they have been trained to closely follow the system prompt & remove the guardrails & bias that was originally trained into the base model.

This makes our models very well suited to Role-Playing games that may require character traits that would be censored by the centralized LLM APIs.

Any game dev will be able to integrate our SDK library into their project to enable NPC's & other aspects of the game to be powered by our small & lightning fast 0.5B – 4B models.

We believe that using our API is better for most gaming use cases compared to running AI models locally on the player’s GPU, since games typically require all available VRAM for graphics.

Synthetic dataset generation suite

Generating datasets with leading closed-source models (eg. ChatGPT & Claude) is extremely expensive. We had a client request an Arabic fine-tuned proprietary model, and they ended up spending around $10k while allowing us to use their GPT-4 API key to generate the synthetic data required for the fine-tune.

Dolphin network will offer a synthetic data generation suite that leverages idle network capacity to reduce per-token costs for non-urgent workloads.

The latest generation of large models that we have trained such as Dolphin Qwen3 235B can generate very high-quality synthetic datasets.

This is especially useful for those who are looking to generate uncensored datasets for model training, fiction or other use cases.

That's something we enable that you simply can't do anywhere else besides Dolphin.

Examples:

• If you want a dataset that takes positions on all sides of an issue.
• You want to generate uncensored data in non-English languages (Dolphin Qwen supports 27 additional languages besides English and Chinese).
• You want to generate conversational data for in-game NPC dialogue in a style that matches the theme of the game to give a unique experience to each player (you could imagine generating conversational data for a barbaric medieval role-playing game… this would likely be limited if you tried to do it through ChatGPT or Claude).

Model creation & fine-tuning suite

A user-friendly interface to submit a training run powered by our network, with support for SFT & RL with the users data or a mix of our data templates.

Users can upload their own datasets, select from our curated options, or combine both.

The primary objective is to provide training infrastructure at a lower cost than major cloud providers, while also serving the no-code market by allowing users to train models on custom data without prior experience.

We would like to replicate an experience similar to Tinker that is powered by the idle compute on Dolphin Network to offer training at substantially lower cost to our competitors.

Dolphin Dev

A data-generation system focused on problem solving with tools inside disposable environments. Most open-source datasets teach pattern matching rather than planning, tool use, verification, and iteration. Dolphin Dev addresses this by deploying LLMs into sandboxed environments to solve challenges with tools, producing datasets that improve our existing model lines and pave the way for an autonomous coding product.

This will make our models performant in high-inference coding applications such as Cline, Cursor, Roo, and similar suites.

With these datasets, we can train dedicated coding models that run autonomously inside worker-deployed containers to complete jobs. A future web interface will allow submitting a job; after completion, the worker returns the work product and logs.

Worker-hosted sandboxes (opt-in)

Nodes on Dolphin Network can dedicate a slice of RAM/CPU to host disposable sandboxes for these jobs. The sandbox and co-located LLM iterate locally - plan → call tools → run tests → refine - so only the final artifact and logs are sent back to the customer. This reduces latency, increases iterations per minute, and monetizes idle CPU/RAM alongside GPU.

Benefits

• Monetize available RAM/CPU as well as GPU by running jobs alongside worker nodes.
• Offer a premium tier for autonomous execution.
• Shift jobs to off-peak hours to flatten demand across the network.

Peer to peer AI development marketplace

A two-sided marketplace for fine-tuning, synthetic data generation, and P2P inference. Each wallet carries a soul-bound reputation NFT that is updated after transactions.

• Roles: Dataset Creator / Dataset Consumer
• Roles: Model Trainer / Model Training Consumer
• Roles: Inference Provider / Inference Consumer

This coordinates peer-to-peer relations across the emerging data and model ecosystem.