Data, models, and compute form the three core pillars of AI infrastructure—comparable to fuel (data), engine (model), and energy (compute)—all indispensable. Much like the evolution of infrastructure in the traditional AI industry, the Crypto AI sector has undergone a similar trajectory. In early 2024, the market was dominated by decentralized GPU projects (such as Akash, Render, and io.net), characterized by a resource-heavy growth model focused on raw compute power. However, by 2025, industry attention has gradually shifted toward the model and data layers, marking a transition from low-level infrastructure competition to more sustainable, application-driven middle-layer development.

Traditional large language models (LLMs) rely heavily on massive datasets and complex distributed training infrastructures, with parameter sizes often ranging from 70B to 500B, and single training runs costing millions of dollars. In contrast, Specialized Language Models (SLMs) adopt a lightweight fine-tuning paradigm that reuses open-source base models like LLaMA, Mistral, or DeepSeek, and combines them with small, high-quality domain-specific datasets and tools like LoRA to quickly build expert models at significantly reduced cost and complexity.

Importantly, SLMs are not integrated back into LLM weights, but rather operate in tandem with LLMs through mechanisms such as Agent-based orchestration, plugin routing, hot-swappable LoRA adapters, and RAG (Retrieval-Augmented Generation) systems. This modular architecture preserves the broad coverage of LLMs while enhancing performance in specialized domains—enabling a highly flexible, composable AI system.

Crypto AI projects, by nature, struggle to directly enhance the core capabilities of LLMs. This is due to:

• High technical barriers: Training foundation models requires massive datasets, compute power, and engineering expertise—capabilities currently held only by major tech players in the U.S. (e.g., OpenAI) and China (e.g., DeepSeek).

• Open-source ecosystem limitations: Although models like LLaMA and Mixtral are open-sourced, critical breakthroughs still rely on off-chain research institutions and proprietary engineering pipelines. On-chain projects have limited influence at the core model layer.

That said, Crypto AI can still create value by fine-tuning SLMs on top of open-source base models and leveraging Web3 primitives like verifiability and token-based incentives. Positioned as the "interface layer" of the AI stack, Crypto AI projects typically contribute in two main areas:

• Trustworthy verification layer: On-chain logging of model generation paths, data contributions, and usage records enhances traceability and tamper-resistance of AI outputs.

• Incentive mechanisms: Native tokens are used to reward data uploads, model calls, and Agent executions—building a positive feedback loop for model training and usage.

The feasible applications of model-centric Crypto AI projects are primarily concentrated in three areas: lightweight fine-tuning of small SLMs, on-chain data integration and verification through RAG architectures, and local deployment and incentivization of edge models. By combining blockchain’s verifiability with token-based incentive mechanisms, Crypto can offer unique value in these medium- and low-resource model scenarios, forming a differentiated advantage in the “interface layer” of the AI stack. An AI blockchain focused on data and models enables transparent, immutable on-chain records of every contribution to data and models, significantly enhancing data credibility and the traceability of model training. Through smart contract mechanisms, it can automatically trigger reward distribution whenever data or models are utilized, converting AI activity into measurable and tradable tokenized value—thus creating a sustainable incentive system. Additionally, community members can participate in decentralized governance by voting on model performance and contributing to rule-setting and iteration using tokens.

OpenLedger is one of the few blockchain AI projects in the current market focused specifically on data and model incentive mechanisms. It is a pioneer of the "Payable AI" concept, aiming to build a fair, transparent, and composable AI execution environment that incentivizes data contributors, model developers, and AI application builders to collaborate on a single platform—and earn on-chain rewards based on actual contributions.

OpenLedger offers a complete end-to-end system—from “data contribution” to “model deployment” to “usage-based revenue sharing.” Its core modules include:

• Model Factory: No-code fine-tuning and deployment of custom models using open-source LLMs with LoRA;

• OpenLoRA: Supports coexistence of thousands of models, dynamically loaded on demand to reduce deployment costs;

• PoA (Proof of Attribution): Tracks usage on-chain to fairly allocate rewards based on contribution;

• Datanets: Structured, community-driven data networks tailored for vertical domains;

• Model Proposal Platform: A composable, callable, and payable on-chain model marketplace.

Together, these modules form a data-driven and composable model infrastructure—laying the foundation for an on-chain agent economy.

On the blockchain side, OpenLedger built on OP Stack + EigenDA, providing a high-performance, low-cost, and verifiable environment for running AI models and smart contracts:

• Built on OP Stack: Leverages the Optimism tech stack for high throughput and low fees;

• Settlement on Ethereum Mainnet: Ensures transaction security and asset integrity;

• EVM-Compatible: Enables fast deployment and scalability for Solidity developers;

• Data availability powered by EigenDA: Reduces storage costs while ensuring verifiable data access.

Compared to general-purpose AI chains like NEAR—which focus on foundational infrastructure, data sovereignty, and the “AI Agents on BOS” framework—OpenLedger is more specialized, aiming to build an AI-dedicated chain centered on data and model-level incentivization. It seeks to make model development and invocation on-chain verifiable, composable, and sustainably monetizable. As a model-centric incentive layer in the Web3 ecosystem, OpenLedger blends HuggingFace-style model hosting, Stripe-like usage-based billing, and Infura-like on-chain composability to advance the vision of “model-as-an-asset.”

ModelFactory is OpenLedger’s integrated fine-tuning platform for large language models (LLMs). Unlike traditional fine-tuning frameworks, it offers a fully graphical, no-code interface, eliminating the need for command-line tools or API integrations. Users can fine-tune models using datasets that have been permissioned and validated through OpenLedger, enabling an end-to-end workflow covering data authorization, model training, and deployment.

Key steps in the workflow include:

• Data Access Control: Users request access to datasets; once approved by data providers, the datasets are automatically connected to the training interface.

• Model Selection & Configuration: Choose from leading LLMs (e.g., LLaMA, Mistral) and configure hyperparameters through the GUI.

• Lightweight Fine-Tuning: Built-in support for LoRA / QLoRA enables efficient training with real-time progress tracking.

• Evaluation & Deployment: Integrated tools allow users to evaluate performance and export models for deployment or ecosystem reuse.

• Interactive Testing Interface: A chat-based UI allows users to test the fine-tuned model directly in Q&A scenarios.

• RAG Attribution: Retrieval-augmented generation outputs include source citations to enhance trust and auditability.

ModelFactory’s architecture comprises six key modules, covering identity verification, data permissioning, model fine-tuning, evaluation, deployment, and RAG-based traceability, delivering a secure, interactive, and monetizable model service platform.

The following is a brief overview of the large language models currently supported by ModelFactory:

• LLaMA Series: One of the most widely adopted open-source base models, known for its strong general performance and vibrant community.

• Mistral: Efficient architecture with excellent inference performance, ideal for flexible deployment in resource-constrained environments.

• Qwen: Developed by Alibaba, excels in Chinese-language tasks and offers strong overall capabilities—an optimal choice for developers in China.

• ChatGLM: Known for outstanding Chinese conversational performance, well-suited for vertical customer service and localized applications.

• Deepseek: Excels in code generation and mathematical reasoning, making it ideal for intelligent development assistants.

• Gemma: A lightweight model released by Google, featuring a clean structure and ease of use—suitable for rapid prototyping and experimentation.

• Falcon: Once a performance benchmark, now more suited for foundational research or comparative testing, though community activity has declined.

• BLOOM: Offers strong multilingual support but relatively weaker inference performance, making it more suitable for language coverage studies.

• GPT-2: A classic early model, now primarily useful for teaching and testing purposes—deployment in production is not recommended.

While OpenLedger’s model lineup does not yet include the latest high-performance MoE models or multimodal architectures, this choice is not outdated. Instead, it reflects a "practicality-first" strategy based on the realities of on-chain deployment—factoring in inference costs, RAG compatibility, LoRA integration, and EVM environment constraints.

Model Factory: A No-Code Toolchain with Built-in Contribution Attribution

As a no-code toolchain, Model Factory integrates a built-in Proof of Attribution mechanism across all models to ensure the rights of data contributors and model developers. It offers low barriers to entry, native monetization paths, and composability, setting itself apart from traditional AI workflows.

• For developers: Provides a complete pipeline from model creation and distribution to revenue generation.

• For the platform: Enables a liquid, composable ecosystem for model assets.

For users: Models and agents can be called and composed like APIs.

LoRA (Low-Rank Adaptation) is an efficient parameter-efficient fine-tuning technique. It works by inserting trainable low-rank matrices into a pre-trained large model without altering the original model weights, significantly reducing both training costs and storage demands.

Traditional large language models (LLMs), such as LLaMA or GPT-3, often contain billions—or even hundreds of billions—of parameters. To adapt these models for specific tasks (e.g., legal Q&A, medical consultations), fine-tuning is required. The key idea of LoRA is to freeze the original model parameters and only train the newly added matrices, making it highly efficient and easy to deploy.

LoRA has become the mainstream fine-tuning approach for Web3-native model deployment and composability due to its lightweight nature and flexible architecture.

Comparison of Fine-Tuning Techniques

OpenLoRA is a lightweight inference framework developed by OpenLedger specifically for multi-model deployment and GPU resource sharing. Its primary goal is to solve the challenges of high deployment costs, poor model reusability, and inefficient GPU usage—making the vision of “Payable AI” practically executable.

OpenLoRA is composed of several modular components that together enable scalable, cost-effective model serving:

• LoRA Adapters Storage: Fine-tuned LoRA adapters are hosted on OpenLedger and loaded on demand. This avoids preloading all models into GPU memory and saves resources.

• Model Hosting & Adapter Merging Layer: All adapters share a common base model. During inference, adapters are dynamically merged, supporting ensemble-style multi-adapter inference for enhanced performance.

• Inference Engine: Implements CUDA-level optimizations including Flash-Attention, Paged Attention, and SGMV to improve efficiency.

• Request Router & Token Streaming: Dynamically routes inference requests to the appropriate adapter and streams tokens using optimized kernels.

The inference process follows a mature and practical pipeline:

1. Base Model Initialization: Base models like LLaMA 3 or Mistral are loaded into GPU memory.

2. Dynamic Adapter Loading: Upon request, specified LoRA adapters are retrieved from Hugging Face, Predibase, or local storage.

3. Merging & Activation: Adapters are merged into the base model in real-time, supporting ensemble execution.

4. Inference Execution & Token Streaming: The merged model generates output with token-level streaming, supported by quantization for both speed and memory efficiency.

5. Resource Release: Adapters are unloaded after execution, freeing memory and allowing efficient rotation of thousands of fine-tuned models on a single GPU.

OpenLoRA achieves superior performance through:

• JIT (Just-In-Time) Adapter Loading to minimize memory usage.

• Tensor Parallelism and Paged Attention for handling longer sequences and concurrent execution.

• Multi-Adapter Merging for composable model fusion.

• Flash Attention, precompiled CUDA kernels, and FP8/INT8 quantization for faster, lower-latency inference.

These optimizations enable high-throughput, low-cost, multi-model inference even in single-GPU environments—particularly well-suited for long-tail models, niche agents, and highly personalized AI.

Unlike traditional LoRA frameworks focused on fine-tuning, OpenLoRA transforms model serving into a Web3-native, monetizable layer, making each model:

• On-chain identifiable (via Model ID)

• Economically incentivized through usage

• Composable into AI agents

• Reward-distributable via the PoA mechanism

This enables every model to be treated as an asset:

In addition, OpenLedger has released future performance benchmarks for OpenLoRA. Compared to traditional full-parameter model deployments, GPU memory usage is significantly reduced to 8–12GB, model switching time is theoretically under 100ms, throughput can reach over 2,000 tokens per second, and latency is kept between 20–50ms.

While these figures are technically achievable, they should be understood as upper-bound projections rather than guaranteed daily performance. In real-world production environments, actual results may vary depending on hardware configurations, scheduling strategies, and task complexity.

High-quality, domain-specific data has become a critical asset for building high-performance models. Datanets serve as OpenLedger’s foundational infrastructure for “data as an asset,” enabling the collection, validation, and distribution of structured datasets within decentralized networks. Each Datanet acts as a domain-specific data warehouse, where contributors upload data that is verified and attributed on-chain. Through transparent permissions and incentivization, Datanets enable trusted, community-driven data curation for model training and fine-tuning.

In contrast to projects like Vana that focus primarily on data ownership, OpenLedger goes beyond data collection by turning data into intelligence. Through its three integrated components—Datanets (collaborative, attributed datasets), Model Factory (no-code fine-tuning tools), and OpenLoRA (trackable, composable model adapters)—OpenLedger extends data’s value across the full model training and on-chain usage cycle. Where Vana emphasizes “who owns the data,” OpenLedger focuses on “how data is trained, invoked, and rewarded,” positioning the two as complementary pillars in the Web3 AI stack: one for ownership assurance, the other for monetization enablement.

Proof of Attribution (PoA) is OpenLedger’s core mechanism for linking contributions to rewards. It cryptographically records every data contribution and model invocation on-chain, ensuring that contributors receive fair compensation whenever their inputs generate value. The process flows as follows:

• Data Submission: Users upload structured, domain-specific datasets and register them on-chain for attribution.

• Impact Assessment: The system evaluates the training relevance of each dataset at the level of each inference, factoring in content quality and contributor reputation.

• Training Verification: Logs track which datasets were actually used in model training, enabling verifiable contribution proof.

• Reward Distribution: Contributors are rewarded in tokens based on the data’s effectiveness and influence on model outputs.

• Quality Governance: Low-quality, spammy, or malicious data is penalized to maintain training integrity.

Compared to Bittensor’s subnet and reputation-based incentive architecture, which broadly incentivizes compute, data, and ranking functions, OpenLedger focuses on value realization at the model layer. PoA isn’t just a rewards mechanism—it is a multi-stage attribution framework for transparency, provenance, and compensation across the full lifecycle: from data to models to agents. It turns every model invocation into a traceable, rewardable event, anchoring a verifiable value trail that aligns incentives across the entire AI asset pipeline.

Retrieval-Augmented Generation (RAG) Attribution

RAG (Retrieval-Augmented Generation) is an AI architecture that enhances the output of language models by retrieving external knowledge—solving the issue of hallucinated or outdated information. OpenLedger introduces RAG Attribution to ensure that any retrieved content used in model generation is traceable, verifiable, and rewarded.

RAG Attribution workflow:

• User Query → Data Retrieval: The AI retrieves relevant data from OpenLedger’s indexed datasets (Datanets).

• Answer Generation with Tracked Usage: Retrieved content is used in the model’s response and logged on-chain.

• Contributor Rewards: Data contributors are rewarded based on the fund and relevance of retrieval.

Transparent Citations: Model outputs include links to the original data sources, enhancing trust and auditability.

In essence, OpenLedger’s RAG Attribution ensures that every AI response is traceable to a verified data source, and contributors are rewarded based on usage frequency, enabling a sustainable incentive loop. This system not only increases output transparency but also lays the groundwork for verifiable, monetizable, and trusted AI infrastructure.

OpenLedger has officially launched its testnet, with the first phase focused on the Data Intelligence Layer—a community-powered internet data repository. This layer aggregates, enhances, classifies, and structures raw data into model-ready intelligence suitable for training domain-specific LLMs on OpenLedger. Community members can run edge nodes on their personal devices to collect and process data. In return, participants earn points based on uptime and task contributions. These points will later be convertible into $OPEN tokens, with the specific conversion mechanism to be announced before the Token Generation Event (TGE).

The second testnet phase introduces Datanets, a whitelist-only contribution system. Participants must pass a pre-assessment to access tasks such as data validation, classification, or annotation. Rewards are based on accuracy, difficulty level, and leaderboard ranking.

OpenLedger aims to close the loop from data acquisition to agent deployment, forming a full-stack decentralized AI value chain:

1. Phase 1: Data Intelligence LayerCommunity nodes harvest and process real-time internet data for structured storage.

2. Phase 2: Community ContributionsCommunity contributes to validation and feedback, forming the “Golden Dataset” for model training.

3. Phase 3: Build Models & ClaimUsers fine-tune and claim ownership of specialized models, enabling monetization and composability.

4. Phase 4: Build AgentsModels can be turned into intelligent on-chain agents, deployed across diverse scenarios and use cases.

Ecosystem Partners: OpenLedger collaborates with leading players across compute, infrastructure, tooling, and AI applications:

• Compute & Hosting: Aethir, Ionet, 0G

• Rollup Infrastructure: AltLayer, Etherfi, EigenLayer AVS

• Tooling & Interoperability: Ambios, Kernel, Web3Auth, Intract

• AI Agents & Model Builders: Giza, Gaib, Exabits, FractionAI, Mira, NetMind

**Brand Momentum via Global Summits:**Over the past year, OpenLedger has hosted DeAI Summits at major Web3 events including Token2049 Singapore, Devcon Thailand, Consensus Hong Kong, and ETH Denver. These gatherings featured top-tier speakers and projects in decentralized AI. As one of the few infra-level teams consistently organizing high-caliber industry events, OpenLedger has significantly boosted its visibility and brand equity within both the developer community and broader Crypto AI ecosystem—laying a solid foundation for future traction and network effects.

OpenLedger completed an $11.2 million seed round in July 2024, Backed by Polychain Capital, Borderless Capital, Finality Capital, Hashkey, prominent angels including Sreeram Kannan (EigenLayer), Balaji Srinivasan, Sandeep (Polygon), Kenny (Manta), Scott (Gitcoin), Ajit Tripathi (Chainyoda), Trevor. The funds will primarily be used to advance the development of OpenLedger’s AI Chain network, including its model incentivization mechanisms, data infrastructure, and the broader rollout of its agent application ecosystem.

Openledger was founded by Ram Kumar, Core Contributor to OpenLedger, who is a San Francisco–based entrepreneur with a strong foundation in both AI/ML and blockchain technologies. He brings a combination of market insight, technical expertise, and strategic leadership to the project. Ram has previously co-led a blockchain and AI/ML R&D company with over $35 million in annual revenue and has played a key role in developing high-impact partnerships, including a strategic joint venture with a Walmart subsidiary. His work focuses on ecosystem development and building alliances that drive real-world adoption across industries.

OPEN is the native utility token of the OpenLedger ecosystem. It underpins the platform’s governance, transaction processing, incentive distribution, and AI Agent operations—forming the economic foundation for sustainable, on-chain circulation of AI models and data assets. Although the tokenomics framework remains in its early stages and is subject to refinement, OpenLedger is approaching its Token Generation Event (TGE) amid growing traction across Asia, Europe, and the Middle East.

Key utilities of OPEN include:

• Governance & Decision-Making:OPEN holders can vote on critical aspects such as model funding, agent management, protocol upgrades, and treasury allocation.

• Gas Token & Fee Payments:OPEN serves as the native gas token for the OpenLedger L2, enabling AI-native customizable fee models and reducing dependency on ETH.

• Attribution-Based Incentives:Developers contributing high-quality datasets, models, or agent services are rewarded in OPEN based on actual usage and impact.

• Cross-Chain Bridging:OPEN supports interoperability between OpenLedger L2 and Ethereum L1, enhancing the portability and composability of models and agents.

• Staking for AI Agents:Operating an AI Agent requires staking OPEN. Underperforming or malicious agents risk slashing, thereby incentivizing high-quality and reliable service delivery.

Unlike many governance models that tie influence strictly to token holdings, OpenLedger introduces a merit-based governance system where voting power is linked to value creation. This design prioritizes contributors who actively build, refine, or utilize models and datasets, rather than passive capital holders. By doing so, OpenLedger ensures long-term sustainability and guards against speculative control—staying aligned with its vision of a transparent, equitable, and community-driven AI economy.

OpenLedger, positioned as a "Payable AI" model incentive infrastructure, aims to provide verifiable, attributable, and sustainable value realization pathways for both data contributors and model developers. By centering on on-chain deployment, usage-based incentives, and modular agent composition, it has built a distinct system architecture that stands out within the Crypto AI sector. While no existing project fully replicates OpenLedger’s end-to-end framework, it shows strong comparability and potential synergy with several representative protocols across areas such as incentive mechanisms, model monetization, and data attribution.

Bittensor is one of the most representative decentralized AI networks, operating a multi-role collaborative system driven by subnets and reputation scoring, with its $TAO token incentivizing model, data, and ranking node participation. In contrast, OpenLedger focuses on revenue sharing through on-chain deployment and model invocation, emphasizing lightweight infrastructure and agent-based coordination. While both share common ground in incentive logic, they differ in system complexity and ecosystem layer: Bittensor aims to be the foundational layer for generalized AI capability, whereas OpenLedger serves as a value realization layer at the application level.

Sentient introduces the “OML (Open, Monetizable, Loyal) AI” concept, emphasizing community-owned models with unique identity and income tracking via Model Fingerprinting. While both projects advocate for contributor recognition, Sentient focuses more on the training and creation stages of models, whereas OpenLedger concentrates on deployment, invocation, and revenue sharing. This makes the two complementary across different stages of the AI value chain—Sentient upstream, OpenLedger downstream.

OpenGradient is focused on building secure inference infrastructure using TEE and zkML, offering decentralized model hosting and trustable execution. It emphasizes foundational infrastructure for secure AI operations. OpenLedger, on the other hand, is built around the post-deployment monetization cycle, combining Model Factory, OpenLoRA, PoA, and Datanets into a complete “train–deploy–invoke–earn” loop. The two operate in different layers of the model lifecycle—OpenGradient on execution integrity, OpenLedger on economic incentivization and composability—with clear potential for synergy.

CrunchDAO focuses on decentralized prediction competitions in finance, rewarding communities based on submitted model performance. While it fits well in vertical applications, it lacks capabilities for model composability and on-chain deployment. OpenLedger offers a unified deployment framework and composable model factory, with broader applicability and native monetization mechanisms—making both platforms complementary in their incentive structures.

Assisterr, built on Solana, encourages the creation of small language models (SLMs) through no-code tools and a $sASRR-based usage reward system. In contrast, OpenLedger places greater emphasis on traceable attribution and revenue loops across data, model, and invocation layers, utilizing its PoA mechanism for fine-grained incentive distribution. Assisterr is better suited for low-barrier community collaboration, whereas OpenLedger targets reusable, composable model infrastructure.

While both OpenLedger and Pond offer “Model Factory” modules, their target users and design philosophies differ significantly. Pond focuses on Graph Neural Network (GNN)-based modeling to analyze on-chain behavior, catering to data scientists and algorithm researchers through a competition-driven development model. In contrast, OpenLedger provides lightweight fine-tuning tools based on language models (e.g., LLaMA, Mistral), designed for developers and non-technical users with a no-code interface. It emphasizes automated on-chain incentive flows and collaborative data-model integration, aiming to build a data-driven AI value network.

Bagel introduces ZKLoRA, a framework for cryptographic verifiable inference using LoRA fine-tuned models and zero-knowledge proofs (ZKP) to ensure the correctness of off-chain execution. Meanwhile, OpenLedger uses LoRA fine-tuning with OpenLoRA to enable scalable deployment and dynamic model invocation. OpenLedger also addresses verifiable inference through a different lens—by attaching proof of attribution to every model output, it explains which data contributed to the inference and how. This enhances transparency, rewards top data contributors, and builds trust in the decision-making process. While Bagel focuses on the integrity of computation, OpenLedger brings accountability and explainability through attribution

Sapien and FractionAI focus on decentralized data labeling, while Vana and Irys specialize in data ownership and sovereignty. OpenLedger, via its Datanets + PoA modules, tracks high-quality data usage and distributes on-chain incentives accordingly. These platforms serve different layers of the data value chain—labeling and rights management upstream, monetization and attribution downstream—making them naturally collaborative rather than competitive.

In summary, OpenLedger occupies a mid-layer position in the current Crypto AI ecosystem as a bridge protocol for on-chain model assetization and incentive-based invocation. It connects upstream training networks and data platforms with downstream Agent layers and end-user applications—becoming a critical infrastructure that links the supply of model value with real-world utilization.

8. Conclusion | From Data to Models—Let AI Earn Too

OpenLedger aims to build a “model-as-asset” infrastructure for the Web3 world. By creating a full-cycle loop of on-chain deployment, usage incentives, ownership attribution, and agent composability, it brings AI models into a truly traceable, monetizable, and collaborative economic system for the first time.

Its technical stack—comprising Model Factory, OpenLoRA, PoA, and Datanets—provides:

• low-barrier training tools for developers,

• fair revenue attribution for data contributors,

• composable model invocation and reward-sharing mechanisms for applications.

This comprehensively activates the long-overlooked ends of the AI value chain: data and models.

Rather than being just another Web3 version of HuggingFace, OpenLedger more closely resembles a hybrid of HuggingFace + Stripe + Infura, offering model hosting, usage-based billing, and programmable on-chain API access. As the trends of data assetization, model autonomy, and agent modularization accelerate, OpenLedger is well-positioned to become a central AI chain under the “Payable AI” paradigm.