What Pluralis Research does

Pluralis Research is a privately held AI research lab developing infrastructure for decentralized, collectively-owned training of foundation models. Founded in late 2023 by Alexander Long, the company is headquartered in San Francisco with operations in the US and Australia, and operates with a small team (1-10 employees) composed entirely of ML PhDs drawn from Google, Anthropic, and Amazon. The company's core offering is Protocol Learning, a methodology and software stack that enables multiple independent participants to collaboratively train and serve large language models over the public Internet using consumer-grade GPUs (16GB+ VRAM for earlier runs, 24GB+ for the current Pluralis-8B run), such that no single participant can ever extract the full model weights. The technical stack includes Subspace Networks (an architectural compressor achieving up to 100x lossless activation compression for WAN-grade pipeline parallelism), AsyncMesh/asynchronous SPARTA for barrier-free within-stage optimization, the Node0 multi-party training library, the Agora platform (a live decentralized pre-training system with a public dashboard), Nesterov-accelerated async pipeline parallelism, and Unextractable Protocol Models that enable fractional model ownership and programmatic value flow from downstream model use back to trainers.

Operationally, Pluralis has validated its thesis at production scale through Node0-7.5B (a completed 7.5B-parameter training run using 1,642 GPUs contributed by 300+ participants from 198 cities, processing 36B tokens over three weeks) and the ongoing Pluralis-8B collective run, which had processed 445.3B tokens with 104 active nodes, 42.08k swarm TFLOPS, and 19.3% MFU at the time of last reporting. The company has accumulated top-tier research output including five NeurIPS 2025 papers (Subspace Networks, Unextractable Protocol Models, Mixtures of Subspaces), ICML 2025 (Nesterov Method for Async PP), and accepted ICML/ICLR 2026 papers (Factored Gossip DiLoCo, Taming Curvature). It has been selected for the AWS 2025 Generative AI Accelerator and runs Protocol Learning workshops at ICML 2026 (Seoul) and Rio de Janeiro.

Pluralis is pre-revenue. The company has not yet established a commercial product or pricing model; participation in training runs is free and permissionless, and the long-term economic mechanism (fractional model ownership flowing value from downstream model use back to contributors) is conceptual rather than implemented. It raised a $7.6M seed round in March 2025 co-led by Union Square Ventures and CoinFund (with Bodhi Ventures, Eden Block, Topology, and Variant), plus up to $1M in AWS credits in October 2025. Its go-to-market is community-led, relying on open-source GitHub releases, technical blog posts, peer-reviewed publications, and academic workshops to attract GPU contributors and research collaborators. Primary customer segments are AI researchers and ML engineers interested in decentralized training, GPU compute contributors who want a stake in resulting models, and open-source AI advocates concerned about power concentration in frontier AI.

Name

Pluralis Research

Legal name

Pluralis Research

Website

https://pluralis.ai

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

1–10 employees

Short description

Pluralis Research is a privately held AI research lab building Protocol Learning, a decentralized framework that lets independent contributors collaboratively train foundation models over the Internet using consumer GPUs, without any participant extracting the full weights. It serves AI researchers, GPU contributors, and open-source AI advocates, and is pre-revenue.

Ownership category

akta.pro rank

Product category

Decentralized AI Training Infrastructure

NAICS

Educational Services (611)

SIC

Services-Prepackaged Software (7372)

akta.pro primary industry

Confidential AI & Privacy-Preserving ML (federated learning, MPC, HE, TEEs) (HDAAAKAI)

Keywords

• Decentralized AI training
• Foundation model development
• Distributed machine learning
• Open source AI infrastructure
• Model parallel training

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Offices2 records

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Infrastructure, Marketing or Sales, Operations

Revenue model

1. Not yet monetizing: Pluralis Research is currently in research phase and has not yet established a revenue model. The company is focused on building the technical infrastructure for decentralized AI training. The long-term vision involves enabling value capture through fractional model ownership via Protocol Models, where downstream use of trained models generates value that flows back to trainers proportional to their contribution.

Go-to-market motion1 record

Distribution channels3 records

Marketing channels6 records

Core offering

Pluralis Research develops Protocol Learning, a decentralized framework for multi-participant collaborative training and inference of foundation models over the internet, where no single participant can ever obtain a full copy of the model. The company runs open, permissionless training runs on the Agora platform that allow anyone with a consumer-grade GPU (16GB+ VRAM) and internet connection to contribute compute and earn fractional ownership in the resulting models, producing open foundation models trained without centralized control.

Product overview

Pluralis Research offers a suite of interconnected products for decentralized, collectively-owned AI training. The core platform is Agora, which enables multi-party pipeline-parallel training over the Internet using consumer hardware. Node0 represents their first public training run demonstrating the technology, while Protocol Learning is the foundational methodology enabling compute pooling with economic sustainability. Protocol Models provide the cryptographic property ensuring models remain unextractable, enabling value capture by trainers. Subspace Networks serve as the critical compression technology enabling low-bandwidth model-parallel training. The Agora Dashboard provides real-time monitoring for active training runs.

Differentiator

Problem solved

Functional benefit

Brands

• Agora: The system that connects consumer GPUs to collaborative training runs, enabling decentralized model-parallel pre-training over the public internet.
• Node0
• Protocol Learning

Products and services

• Agora Platform A decentralized pipeline-parallel pre-training system that connects consumer GPUs to collaborative training runs, enabling multi-party foundation model training where participants host individual model stages over the public internet. Includes live swarm participation, state synchronization, and a public dashboard with real-time metrics.
• Node0 Training Run The first permissionless, multi-participant model-parallel pretraining run from Pluralis Research, enabling anyone with a single 16GB+ consumer GPU to join collaborative training of models larger than any individual could train alone, operating over standard internet connections.
• Protocol Learning The foundational research paradigm enabling collectively-owned AI through trustless pooling of compute from multiple participants for collaborative model training and hosting within protocols, with economic sustainability ensured by Unextractable Protocol Models that prevent any single party from capturing full model value.
• Pluralis-8B Collective Run Current live collaborative training run for an 8B-parameter transformer model, with participation open via the Agora system to anyone with a 24GB+ GPU (RTX 4090, RTX 5090, RTX 6000), 80GB RAM, and 200 Mbps network connection.

Quantifiable outcome

• Up to 100x compression in inter-stage communication via Subspace Networks, enabling training over 80-200 Mbps networks
• +3 more outcomes

Segments3 records

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability5 records

Feature6 records

Partnerships

Three partnerships are on record, tiered core and minor.

• Amazon Web Services (AWS)coreStrategic or Co-development Partner · 1 October 2025Pluralis Research was selected for the AWS 2025 Generative AI Accelerator program, receiving up to $1 million in AWS credits along with technical mentorship and access to the AWS generative AI technology stack. The eight-week program runs from Amazon's headquarters in Seattle to AWS re:Invent in Las Vegas, supporting machine learning performance, stack optimization, and go-to-market strategy.
• Prime IntellectminorStrategic or Co-development Partner · 1 March 2024Prime Intellect is a decentralized training network with an ERC-20 token, also backed by CoinFund ($5.5M seed + $15M follow-on). Both companies are part of the emerging decentralized AI ecosystem working on distributed GPU resources and permissionless foundation model training. Mentioned as a portfolio company alongside Pluralis Research.
• GensynminorStrategic or Co-development Partner · 1 March 2024Gensyn is a decentralized training network with an ERC-20 token, backed by CoinFund alongside Pluralis Research. Both companies are building infrastructure for permissionless, globally distributed foundation model training as part of the decentralized AI movement.

Scale indicators9 records

Recent moves6 records

Expansion highlights6 records

Direct peers

• Gensyn: Decentralized machine learning compute network with ERC-20 token, co-backed by CoinFund alongside Pluralis. Both target permissionless foundation model training using distributed GPUs and share the same Web3-aligned investor base and value capture thesis.
• Prime Intellect: Decentralized AI training network with ERC-20 token, also backed by CoinFund. Both companies pursue permissionless, globally distributed foundation model training with token-based contributor incentives, making Prime Intellect Pluralis's closest head-to-head competitor.
• EleutherAI: Open-source AI research collective that pioneered open foundation model training (GPT-Neo, GPT-J, Pythia). Shares Pluralis's mission of democratizing foundation model training and publishes research openly at the model layer.
• Nous Research: Open-source AI research lab focused on decentralized, collectively-trained foundation models (e.g., Hermes, FineTuning recipes). Shares Pluralis's open-research ethos, decentralized training philosophy, and target user base of independent ML researchers.
• Bagel: Decentralized AI/ML network for collaborative model training and inference with crypto-economic incentives. Competes with Pluralis for the same permissionless, contributor-paid training model.

Emerging players

• FedML: Federated learning and collaborative AI training platform supporting distributed training across heterogeneous devices. Shares Pluralis's interest in multi-party, low-bandwidth distributed training but focuses on federated rather than pipeline-parallel approaches.
• Petals (BigScience): Community-run system for collaborative LLM inference and training using consumer GPUs. Closely aligned with Pluralis's Agora platform concept of pipeline-parallel model hosting across volunteer consumer hardware.

Broad incumbents

• Together AI: Decentralized AI cloud platform offering open-source model hosting, training, and inference at scale. While commercially focused (unlike Pluralis's research orientation), it competes for the same decentralized AI infrastructure narrative and contributor/developer mindshare.
• Hugging Face: Central platform for open-source AI model sharing, training, and collaboration. While not decentralized training-native, Hugging Face is the dominant open-source AI community hub and competes for the same developer/researcher attention Pluralis targets.

Others

• Determined AI: Open-source distributed training platform (acquired by HPE) focused on efficient multi-GPU/multi-node training. Provides adjacent infrastructure that could complement or compete with Pluralis's distributed training stack for non-decentralized use cases.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks7 records

Key highlights6 records

Customer concentration

• Xx.com/PluralisHQ
• GitHubgithub.com/PluralisResearch

Revenue estimate

Valuation estimate

Number of profiles

Profiles2 records

Funding overview

Funding rounds3 records

Investors8 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Pluralis Research