Bagel (I): Monetizing Open Source AI

No More Piracy. Here's how Bagel's ZKLoRA is changing the game

Want More? Follow Teng Yan and Chain of Thought on X

Join our Telegram research channel and Discord community.

Subscribe for timely research and insights—delivered to your inbox.

TL;DR

  • Bagel is tackling one of open-source AI’s biggest challenges: monetization.

  • By combining advanced cryptography with privacy-preserving machine learning, Bagel enables open-source AI model contributors to generate value without sacrificing transparency.

  • ZKLoRA is a breakthrough that allows developers to share and protect their fine-tuned AI adapters using zero-knowledge proofs (ZKPs). It ensures fair attribution and monetization while keeping proprietary details private.

  • Because ZKLoRA makes verification fast and cheap, it democratizes access to fine-tuned AI. Even on massive models like Llama-70B with 80 LoRA modules, verification remains practical and efficient (< 2 minutes).

Let’s start with a thought experiment.

You’ve built an extraordinary open-source AI model capable of interpreting MRI scans more accurately than any radiologist. It’s a groundbreaking tool that could diagnose cancer early and save thousands of lives.

But there’s a catch—the journey to get there was grueling. Months of your life are spent battling internal departments for data access. Countless late nights painstakingly labeling MRI images after a full day of work. Tens of thousands of dollars sunk into GPUs, possible only from hard-fought grant funding.

It’s exhausting, but you tell yourself it’s worth it. This model will change the game. It will save lives. It’s a pain I know all too well, having spent half my career in healthcare.

But then, one morning, you wake up to your worst nightmare. Your open-source masterpiece, which you poured blood, sweat, and grants into, has been cloned. Some developer, who probably didn’t spend a second labeling data, is now selling it as a premium product. And you? You don’t even get a thank-you note.

Open Source Is Broken

Unfortunately, this is the everyday story of open-source AI. The reality for many creators.

Open-source AI has democratized innovation, putting powerful models into the hands of anyone with a laptop and an internet connection. Collaboration, open datasets, and transparent codebases have fuelled breakthroughs that would’ve been impossible in a world dominated by proprietary systems.

But there’s a glaring issue: the people powering this ecosystem—developers, researchers, and data contributors—rarely see any recognition or compensation. Monetization feels like an afterthought, left to corporations that build profitable layers on top.

And without good monetization models, you can only rely on the goodwill of others to make progress. What happens if Zuck stops making the next generation of Llama models open source? What if there was no DeepSeek?

Let's be honest — we’re screwed.

So, what do we do about it?

How do we keep open-source AI vibrant, fair, and sustainable?

This is the story of Bagel.

Bagel is taking on one of open-source AI’s biggest challenges: monetization. By combining advanced cryptography with privacy-preserving machine learning, it’s building a platform where open AI models can generate value without sacrificing transparency. If it works, it could reshape the economics of open-source innovation.

Our deep dive unpacks Bagel’s vision. Part one explores its origins and cryptographic breakthroughs with ZKLoRA. Part two breaks down the “Bakery” protocol, its actual traction thus far, and how it tackles the monetization gap practically.

If Bagel succeeds, it could be the lifeline open-source AI desperately needs. Let’s dive in.

Innovation vs Exploitation

Open-source AI thrives on transparency and collaboration.

It’s a system where anyone can inspect, refine, and build on existing models, fostering collective innovation. But openness comes with a cost—exploitation.

The moment a model is public, it’s a free game. Cloning, rebranding, and profit-taking without attribution run rampant leaving original creators without recognition or reward.

The Open Source Initiative (OSI) is working to define open-source AI formally, ensuring access to data, source code, and model parameters. But a definition alone won’t solve the biggest issue: sustainability. Developers have tried everything—donations, premium versions, SaaS—but none have cracked the problem of fair attribution and funding.

I’m heartened that there are still success cases.

Let’s go back to healthcare as an example.

AI has incredible potential to revolutionize medicine, yet progress has been frustratingly slow. Why? Because so much of the field is stuck in closed-loop systems. Hospitals, researchers, and corporations hoard their data, and even open-source projects face the uphill battle of scaling without a sustainable model..

One notable exception is MONAI.

Launched in 2019 by NVIDIA, NIH, and King’s College London, it’s an open-source deep-learning framework tailored for medical imaging. With powerful tools like MAISI, a 3D diffusion model that generates synthetic medical images, MONAI has revolutionized AI research without relying on sensitive patient data.

But MONAI’s success exposes a hard truth: even the best open-source projects need deep-pocketed backers. NVIDIA funds MONAI’s growth through enterprise licensing, hardware sales, and platform integrations. Without this, even a groundbreaking project like MONAI could struggle to survive.

So here’s the real question: How do we create more MONAIs—AI projects that remain open while sustaining themselves independently?

If open-source AI is to reach its full potential, we need a model that protects creators, ensures attribution, and makes transparency profitable, not just idealistic.

Closed AI Is Raking in the Cash

Meanwhile, the AI landscape is dominated by closed-source giants like Anthropic, OpenAI, and now the $500 billion Stargate project, which centralizes power and resources in ways that could stifle grassroots innovation.

And they’re raking in the revenue.

In 2024, OpenAI’s API alone is generating an estimated $500M - $800M each year, with top-line revenue growing at a whopping 248% annually. The profit margins on its API are estimated to be at least 50%, with the most expensive o1 model likely boasting margins of 70 - 80%.

If you minus the spending on training AI and just focus on the inference, they’ve mastered the art of turning AI into a lucrative business—while open-source developers scrape together resources for their next GPU.

For many, this imbalance feels both unfair and inevitable. Open source was never designed to make its creators rich. Just look at Linux. It runs 60% of the world’s servers, yet Linus Torvalds never cashed in billions from selling it.

The Lightbulb Moment

Bagel’s founder, Bidhan Roy, saw the problem up close. Managing machine learning infrastructure for Amazon Alexa, where 100M+ devices ran on his systems, he knew open AI (pun intended) could scale.

The problem wasn’t capability—it was funding. Without deep pockets for training, open models were either ignored or stolen by those who could afford to deploy them better.

In 2017, while working on Bitcoin trading at Cash App, Bidhan had a realization: cryptography could fix this. What if open-source AI had built-in mechanisms to recognize and reward contributors automatically?

The idea was bold but doable—a system where anyone sharing data, compute, or model training could get compensated fairly and transparently.

That’s how Bagel was born.

Bagel: Attribution + Privacy + Ownership

If I had to sum it up,” says Bidhan, “we’re making open-source AI monetisable using cryptography.”

In practice, that means three pillars:

  1. Attribution: Every contributor to improving the model—whether adding code or data—gets credited. When revenue flows, cryptographic protocols ensure fair, automatic distribution.

Simple, right? Not quite. Attribution is messy.

Where does value actually accrue? How do you measure a dataset’s impact on model performance? One promising approach focuses on benchmark improvements—if a model gets better on these benchmarks after new data is added, that contributor deserves a cut.

  1. Privacy: Bagel’s ecosystem supports privacy-preserving machine learning (PPML), so you can share your data or model parameters without actually revealing them.

  2. Ownership: Instead of conventional licensing agreements, contributors maintain perpetual ownership of their innovations. How? Through privacy-preserving containers and parameter obfuscation. Your work stays yours, locked up in a digital vault, ensuring you always have a stake in what you help build.

The result is the trifecta for open-source AI: preserve privacy, ensure fair attribution, and retain ownership—while getting paid.

For open-source AI developers, it’s like finally getting compensated for picking the low-hanging fruit they’ve been growing all along.

ZK that actually works

Bagel’s latest research on zero-knowledge Low-Rank Adaptation (ZKLoRA) marks a breakthrough in collaboration, monetization, and trust in open-source AI. They have also open-sourced the code for all to use. (go Bagel!)

With ZKLoRA, developers no longer have to choose between sharing and protecting their work. They can have both.

There’s a lot to unpack, so we’ll break it down step by step so you get a clear understanding by the end of this essay.

Why Fine-Tuning Matters

Let’s start with a quick primer on fine-tuning AI models.

In the AI world, the focus is shifting from pre-training (where only a handful of companies can truly compete), to post-training and fine-tuning, where the productivity gains are just beginning

Consider an open-source base model like Llama-3. It’s powerful and great for broad tasks, but not particularly useful when you need it to do something highly specific, like legal document analysis or medical diagnosis.

So we need to fine-tune these models: take the base model, feed it proprietary data, and create a tailored version for your use case.

LoRA (Low-Rank Adaptation) is a lightweight, parameter-efficient fine-tuning method. Which means only the required parts of a large model are fine-tuned while keeping the base model frozen.

Instead of retraining the entire model, LoRA trains a compact adapter module that layers on top of the frozen base model. These adapters are highly tuned towards a specific task. It’s fast, cheap, and efficient.

This architecture works like Lego—millions of developers can stack, mix, and build on each other’s modular contributions.

Compared to Retrieval-Augmented Generation (RAG)—which constantly queries a database (adding latency) or stuff data into a context window (driving up token costs)—LoRA is a one-time investment. Once fine-tuned, the model can run indefinitely without extra compute costs, making it a scalable, cost-efficient alternative.

Enter ZKLoRA

LoRA is a great way to fine-tune AI models, but it comes with a major dilemma—if you share your adapter openly, how do you stop others from copying it? And if you sell it, how does the buyer know it even works?

With ZKLoRA, Bagel solves both problems by integrating zero-knowledge proofs (ZKPs). ZKPs are like showing someone the answer to a puzzle without revealing how you solved it.

In LoRA’s case, ZKLoRA lets developers prove their fine-tuned adapter works with a base model—without ever exposing proprietary details. 

Here’s what that means in practice:

 For developers: You can monetize your fine-tuned adapters without fear of being copied. Your secret sauce stays secret. The source code does not have to be open for the system to function well.

 For buyers: You get proof that the adapter does what it claims without taking the seller’s word for it. No guesswork, no blind trust.

The magic of ZKLoRA lies in its streamlined, three-step workflow:

  1. Multi-Party Inference

The base model owner and the LoRA creator exchange encrypted activations—like swapping puzzle pieces without revealing the full picture.

  • The base model processes part of the input locally, generating partial results (activations)

  • These activations are sent to the LoRA contributor, who uses their private LoRA weights to compute adjustments

This ensures that sensitive data—both the base model parameters and LoRA weights—remain private throughout the interaction.

  1. Proof Generation:

Now, the LoRA contributor has made adjustments, but the base model owner needs proof that these changes are valid—without revealing the actual LoRA model.

To solve this, the contributor generates cryptographic proofs, verifying that their transformations follow the LoRA methodology. These proofs confirm the adapter’s compatibility with the base model without ever exposing proprietary LoRA weights.

In short, the contributor is saying, “Trust me, this works,”—and the math proves it.

  1. Fast Verification:

Finally, the base model owner quickly verifies the proofs, taking only 1–2 seconds per module.

 If any proof fails, the corresponding LoRA module is rejected.

 If all proofs pass, the process continues seamlessly.

The speed of verification allows models to stack multiple LoRA modules without slowing down significantly.

So ZKLoRA is fast and scalable.

Even on massive models like Llama-70B with 80 LoRA modules, verification remains practical and efficient (< 2 minutes).

This is a huge breakthrough, considering that zero-knowledge proofs (ZKPs) have historically been slow and computationally expensive. ZKLoRA somehow makes zero-knowledge cryptography look… efficient.

How was Bagel able to solve this?

Early ZK approaches in machine learning focused on proving compute usage—a flawed metric. It checks that a set of mathematical operations was performed without bothering with the output quality.

In reality, in training, tons of compute goes to waste (overfitting, bad hyperparameters etc.) when results don’t meet evaluation standards. The real question is not compute usage, but whether a contribution actually improves the model.

Second, Bagel keeps it modular instead of verifying entire models with billions of parameters. By only verifying the LoRAs, it makes the process lightweight. Bagel focuses on optimizing verification to prove impact, not effort—with ultra-low latency.

ZKLoRA is a Big Deal

With this, we can imagine a marketplace for fine-tuned AI adapters—legal, medical, gaming, you name it. Developers can sell verified, plug-and-play modules without worrying about piracy. Buyers, in turn, get cryptographic proof that what they’re buying works. Everyone wins.

Open-source AI thrives on global collaboration, but working with anonymous or remote contributors has always been a risky endeavor. ZKLoRA eliminates the need for blind trust.

You don’t need to trust the person—you just need to trust the math. And math is never wrong.

Because ZKLoRA makes verification fast and cheap, it democratizes access to fine-tuned AI. Small developers with limited resources can now confidently participate in the ecosystem, creating a broader, more diverse range of fine-tuned solutions.

By protecting their work, we can create the right incentives for developers to build specialized adapters.

Closing Thoughts

Bagel is rewriting the rules of open-source AI.

By fusing cryptography with privacy-preserving machine learning, it’s turning open-source into a sustainable, monetizable ecosystem—without killing its collaborative spirit. Creators can finally get paid.

If it works, open-source AI might finally have the business model to stand toe-to-toe with the closed-source giants. And that’s a future worth betting on

Cheers,

Teng Yan

In Part two, we’ll break down the “Bakery” protocol, Bagel’s open source AI marketplace, and how it addresses ther monetization gap. Plus: how you can get involved.

You can follow Bagel on X to stay updated on the latest developments

This research series is done in collaboration with Bagel, with Chain of Thought receiving funding for this initiative. All insights and analysis are our own. We uphold strict standards of objectivity in all our viewpoints.

To learn more about our approach to sponsored Deep Dives, please see our note here.

This report is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investment choices.

Reply

or to participate.