Pluralis Research
Decentralized model-parallel training protocol enabling low-bandwidth collaboration and unmaterializable models
Performances
Comparison
Details
This content is generated by RD AI and is for reference only
Pluralis is a decentralized AI training protocol that shards models across globally heterogeneous nodes, enabling collaborative model-parallel training and inference under low bandwidth. Its core mechanism is subspace networks compressing activation transfers, paired with an unmaterializable model design that prevents any single participant from obtaining full weights.
Fundraising
Follow Updates
Follow Lists
People
About Pluralis Research
Pluralis Research is building a decentralized protocol for training large AI models. Its core business is Protocol Learning: a method that splits a model across many independent participants who train it collaboratively over the internet. Instead of sharing full model weights, nodes exchange only small gradient or activation tensors, making it impossible for any single party to extract or copy the complete model. This "unmaterializable model" property ensures that no participant can run away with the final weights, while still allowing open contribution and collective ownership.
The main pain point it addresses is the centralization of AI development. Training frontier-scale models normally requires massive, concentrated compute clusters, which excludes most individuals and small organizations. Existing decentralized approaches either demand high-bandwidth connections (data parallelism) or fail to protect the trained model from being stolen. Pluralis solves both by enabling low-bandwidth model-parallel training and by making the model inherently non-excludable yet non-copyable, creating economic incentives for contributors without needing a trusted coordinator.
In the past six months, Pluralis has moved from theory to practice. It ran a public training experiment using 14 Apple Silicon Macs distributed across four countries (Paris, Zürich, Dublin, Toronto), training a 7.5B-parameter model for three weeks on 36B tokens. The final loss reached 2.75, comparable to a centralized baseline. The run involved 303 active nodes, with 78 contributors each providing over 1 exaFLOP of compute, despite many operating on sub-100Mbps connections. This demonstrated that heterogeneous, unreliable, low-bandwidth consumer hardware can indeed train competitive models. The team also published detailed technical write-ups and opened a GitHub repository, though the project remains at proof-of-concept stage with no production deployment yet.
Pluralis Research was founded in January 2024 by Alexander Long, a former FAANG researcher, with the goal of developing true open-source AI through decentralized training. In March 2025, the company announced a $7.6 million seed round co-led by USV and CoinFund, with participation from Balaji Srinivasan and HuggingFace co-founder Clem Delangue. The company pioneered Protocol Learning, a method allowing models to be trained across distributed networks without any single party holding full weights. A key milestone was a public training run of a 7.5B-parameter OLMo-style model over the internet using 300+ globally distributed consumer GPUs for over three weeks, demonstrating the feasibility of decentralized training at scale. This validated their approach and laid groundwork for future open model development.
Pluralis Research has made significant strides in decentralized AI over the past six months. In March 2026, it unveiled Agora, a system enabling permissionless, internet-scale pretraining of large models. This culminated in Pluralis-8B, an 8.6B-parameter model trained on 500B tokens over 40 days using 330 heterogeneous consumer GPUs, achieving 63% of a centralized H100 baseline's efficiency. In June, the team introduced Stoa, a decoupled RL post-training framework that leveraged 14 consumer Macs across four countries to fine-tune LFM2.5-8B-A1B, boosting pass@1 on PaperSearchQA from 29% to 63%. The company also hosted an ICML workshop on Protocol Learning, formalizing its vision of low-bandwidth, heterogeneous, multi-party model training with economic incentives. Looking ahead, Pluralis plans to integrate Agora and Stoa into a unified RL training loop, tackle open problems in asynchronous optimization and Byzantine robustness, and launch its first fully open protocol-based training run. With a $7.6M seed round led by USV and CoinFund, it is expanding its team and compute partnerships to scale toward frontier-level open models.
Co-founders and core executives:
- Alexander Long (Founder)
- Gil Avraham
- Yan Zuo
- Sameera Ramasinghe
- Ajanthan Thalaiyasingam