Prime Intellect
Last reviewed
Jun 8, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 ยท 1,909 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 ยท 1,909 words
Add missing citations, update stale details, or suggest a clearer explanation.
Prime Intellect is a San Francisco based artificial intelligence company building infrastructure for decentralized AI development, with a focus on training large models across heterogeneous, geographically distributed GPU clusters rather than inside a single data center. The company operates a compute marketplace, publishes open-source frameworks for fault-tolerant distributed training and reinforcement learning, and releases the open-weight INTELLECT series of models. Its stated mission is to build "the full stack for open superintelligence," positioning open, decentralized model development as an alternative to the centralized frontier laboratories that dominate the field.[1][2]
Prime Intellect rose to prominence by demonstrating that frontier-scale training, long assumed to require tightly coupled supercomputers, can be carried out over the public internet across multiple continents. Its 2024 release of INTELLECT-1 was described as the first globally distributed training run of a 10-billion-parameter model, and subsequent releases extended the approach to reinforcement learning at the 32-billion-parameter scale and to a 106-billion-parameter mixture of experts model.[3][4][5]
Prime Intellect's product is a vertically integrated stack for open AI development. At the base is a compute marketplace, sometimes called the Compute Exchange, that aggregates GPU capacity from cloud providers, data centers, and individual contributors and connects it with developers who want to rent it. On top of that sit open-source training frameworks, a hub of reinforcement-learning environments and evaluations, and the company's own open-weight INTELLECT models, which serve both as research demonstrations and as flagship artifacts of the platform.[1][2]
The unifying thesis is that the resources needed to build capable models, namely compute, data, and training recipes, should not be locked inside a handful of well-funded labs. By making distributed training practical across loosely connected, heterogeneous hardware, Prime Intellect argues that a global community can pool idle or surplus compute to train competitive models collaboratively. The company open-sources its frameworks, model weights, training code, and datasets to support this vision.[1][3]
Prime Intellect was founded in 2023 by Vincent Weisser, who serves as chief executive officer, and Johannes Hagemann. The company is headquartered in San Francisco, California. Weisser had previously been active in decentralized science and Web3 communities, while Hagemann had worked as an AI research engineer at the German foundation-model company Aleph Alpha before co-founding Prime Intellect.[6][7]
The company emerged publicly in April 2024 alongside its first announced financing. From the outset it framed itself around the idea of "commoditizing intelligence" by aggregating global compute and enabling collaborative, open-source model training, rather than building proprietary models behind a closed API.[6][8]
The central technical challenge Prime Intellect addresses is that conventional large-scale training relies on very high-bandwidth interconnects between accelerators, which is feasible only within a single data center. Training across geographically separated machines connected by ordinary internet links requires dramatically reducing how often nodes must communicate, while tolerating slow, unreliable, and untrusted participants.[3][9]
In July 2024, Prime Intellect released OpenDiLoCo, an open-source implementation and scaling of DiLoCo (Distributed Low-Communication training), a method originally introduced by Google DeepMind. DiLoCo lets worker groups train locally for many steps and synchronize only occasionally, reducing inter-node communication by roughly a factor of 500 compared with standard data-parallel training. Built on top of the Hivemind library, OpenDiLoCo reproduced DeepMind's results, scaled the method to roughly three times the original model size, and demonstrated training across two continents and three countries while maintaining 90 to 95 percent compute utilization.[9]
For its 10-billion-parameter run, Prime Intellect developed the PRIME framework, a fault-tolerant distributed training system that allows nodes to join and leave dynamically, handles failures gracefully, and combines DiLoCo with a custom int8 all-reduce to compress synchronization traffic. Together these techniques achieved an approximately 400-fold reduction in communication bandwidth relative to standard data-parallel training at the 10-billion-parameter scale.[3]
The company later built PRIME-RL, a framework purpose-built for fully asynchronous, distributed reinforcement learning. To make RL work across untrusted, permissionless contributors, Prime Intellect introduced supporting components including TOPLOC, a locality-sensitive hashing scheme for verifying that inference rollouts from untrusted workers are genuine, and SHARDCAST, a system for efficiently broadcasting updated model weights from training nodes to inference workers. These tools target the trust and bandwidth problems that arise when reinforcement learning is spread across a heterogeneous swarm.[4][5]
In 2025 Prime Intellect launched the Environments Hub, a community platform for aggregating and sharing reinforcement-learning environments and evaluations, paired with an open-source library called verifiers for building them. Environments define tasks and scoring functions that plug directly into the PRIME-RL training loop, and the hub is designed to crowdsource the large library of environments needed to train capable agentic models. Contributors have included organizations such as Arcee AI and Groq alongside many individual developers.[10]
Prime Intellect's INTELLECT series functions as the public proof of its distributed methods. The table below summarizes the main releases.
| Model | Size | Released | Key characteristic |
|---|---|---|---|
| INTELLECT-1 | 10B (dense) | November 2024 | First globally distributed 10B-parameter pretraining run |
| INTELLECT-2 | 32B (dense) | May 2025 | First 32B model trained via globally distributed reinforcement learning |
| INTELLECT-3 | 106B mixture of experts (12B active) | November 2025 | Largest open INTELLECT model; SFT plus large-scale RL on a mixture-of-experts base |
INTELLECT-1, released on November 29, 2024, was a 10-billion-parameter language model trained on roughly one trillion tokens. The run used up to 112 H100 GPUs spread across as many as eight non-colocated data centers in five countries on three continents, achieving about 83 percent overall compute utilization across continents and up to 96 percent when nodes were confined to the United States. Prime Intellect released both base and instruction-tuned checkpoints along with the training code and a technical report.[3][11]
INTELLECT-2, released on May 11, 2025, was a 32-billion-parameter reasoning model post-trained from Alibaba's QwQ-32B using globally distributed, asynchronous reinforcement learning across a permissionless swarm of compute contributors. Released under the Apache 2.0 license with weights, code, and training logs, it improved on QwQ-32B on mathematics and coding benchmarks. The company emphasized that the principal achievement was methodological, showing that reinforcement learning can be conducted over distributed infrastructure, rather than a dramatic capability jump over an already heavily RL-trained base model.[4][12]
INTELLECT-3, released on November 26, 2025, is a 106-billion-parameter mixture-of-experts model with about 12 billion active parameters per forward pass. It was post-trained from the GLM-4.5-Air-Base model using supervised fine-tuning followed by large-scale reinforcement learning. Notably, both stages, including ablations, were carried out on a single colocated cluster of 512 NVIDIA H200 GPUs across 64 nodes over roughly two months, rather than on a globally distributed swarm; the distributed contribution came largely through the crowdsourced RL environments that fed training. Prime Intellect reported strong reasoning results for the model's size, including 98.1 percent on MATH-500, 90.8 percent on AIME 2024, 74.4 percent on GPQA Diamond, and 69.3 percent on LiveCodeBench v6, and open-sourced the full recipe including weights, the PRIME-RL framework, the verifiers library, datasets, and environments.[5][13]
Beyond the numbered models, Prime Intellect has released or supported related artifacts such as the INTELLECT-MATH reasoning model and METAGENE-1, a foundation model for genomic and pandemic-monitoring applications developed with academic collaborators.[1]
Prime Intellect has raised capital across several rounds, with funding databases and the company's own announcements differing slightly on exact figures. The funding figures below should be read as approximate and are attributed to their sources.[1][6][14]
| Round | Date | Amount | Lead investor(s) |
|---|---|---|---|
| Seed | April 2024 | ~$5.5M | Distributed Global, CoinFund |
| Seed extension / Series A | February 2025 | ~$15M | Founders Fund |
| Series B | December 2025 | ~$49.9M (reported) | Founders Fund and others |
The April 2024 seed round of about $5.5 million was co-led by the crypto-focused investors Distributed Global and CoinFund, with participation from Compound, Collab+Currency, and Protocol Labs founder Juan Benet.[6][8]
In late February 2025 the company announced a roughly $15 million round, framed as building its "open superintelligence stack," led by Founders Fund with participation from Menlo Ventures and a roster of prominent angel investors including Andrej Karpathy, Hugging Face chief executive Clement Delangue, Balaji Srinivasan, Tri Dao, Emad Mostaque, and others. Some sources describe this as a Series A and others as a seed extension; the company itself stated the round brought total funding to over $20 million.[1][14]
Funding databases including Crunchbase and Tracxn report a substantially larger Series B of approximately $49.9 million closing around December 2025, with Founders Fund again among the investors, bringing reported cumulative funding to roughly $70 million. As of mid-2026, the company had not published a single canonical blog post detailing this later round, and reports of a financing in the $60 million range around early 2026 should be treated as unconfirmed pending primary disclosure. Readers should therefore weight database figures for the later rounds accordingly.[14][15]
Prime Intellect is one of the most visible proponents of decentralized training as a credible alternative to the centralized scaling paradigm. If models comparable to those from major labs can be trained across pooled, geographically dispersed, and partly volunteer compute, the economic and political dynamics of frontier AI could shift, lowering barriers to entry and reducing dependence on a small number of hyperscale data centers. The company's open releases of weights, code, and training recipes also feed directly into the broader open-source model ecosystem.[2][3]
The approach carries real caveats. The INTELLECT models have so far been post-trained from strong existing open-weight bases such as QwQ-32B and GLM-4.5-Air rather than pretrained from scratch at frontier scale, and INTELLECT-3's heaviest training stages ran on a conventional colocated H200 cluster. Commentators have noted this tension between the decentralized vision and the centralized reality of the largest runs, viewing Prime Intellect's work as an important step toward, rather than a completed demonstration of, fully decentralized frontier training.[5][13]
In the competitive landscape, Prime Intellect sits alongside other organizations pursuing distributed or decentralized training, including Nous Research, Pluralis Research, and Gensyn, while competing against the centralized frontier laboratories whose closed models it positions itself against. Its combination of a compute marketplace, open frameworks, a crowdsourced RL environment hub, and headline open models distinguishes its full-stack strategy within that group.[2][10]