Prime Intellect

Introducing PCCL, the Prime Collective Communications Library — a low-level communication library built for decentralized training across the globe. After six months of dedicated research led by Michael Keiblinger, we are releasing PCCL to accelerate research on distributed low-communication optimization algorithms and further close the gap to centralized training. PCCL has been designed from the ground up for fault tolerance to cleanly and quickly recover from any possible IO failure. Communication libraries like Nvidia's NCCL assume fast, stable connections—ideal for supercomputers, not the public internet. In our testing PCCL achieves up to 45 Gbit/s of bandwidth across data centers in Western Europe and 25 GBit/s training intercontinental across North America and Western Europe. If you want to see how we use PCCL, refer to our open source prime repo, which is our production-ready open source implementation of DiLoCo. Detailed Technical Report: https://xmrrwallet.com/cmx.plnkd.in/gyCKwz6M Github: https://xmrrwallet.com/cmx.plnkd.in/g8hainew Docs: https://xmrrwallet.com/cmx.plnkd.in/gy6baCZt Blog: https://xmrrwallet.com/cmx.plnkd.in/gJsuWxfW

Prime Intellect Meetup July 18th Berlin Join us https://xmrrwallet.com/cmx.plu.ma/m86mfwof

We asked William Brown (Research Lead at Prime Intellect) about Zuckerberg's latest recruitment strategy. "The goal is to bridge the gap from research to product." "Meta's strength is having very good academic researchers that write important papers, but it's the kind of stuff that doesn’t really turn into exciting products." "It's more like betting 5-10 years out versus going through a pre-training cycle for a Gemini, Claude, or OpenAI model." "They want like 10-15 great people who have been through research-to-product cycles." You can listen to the full conversation with Will here: https://xmrrwallet.com/cmx.plnkd.in/gBtAAWxn Follow TBPN for more.

Excellent work by Prime Intellect's Leonardo Federico on the SYNTHETIC-2 launch video. SYNTHETIC-2 supports heterogeneous compute with TOPLOC v2 verifiable proofs, ensuring honest competition for all contributors doing meaningful work towards frontier-level AGI research. No waitlist, no permissions, no compute caps—just spin up your GPUs and start helping us advance towards open-source superintelligence.

Today, we are launching SYNTHETIC-2 Our next-generation, open-source reasoning dataset and planetary-scale pipeline-parallel inference run. Built on our peer-to-peer inference stack and powered by the new DeepSeek-R1-0528 model, SYNTHETIC-2 generates verified reasoning traces spanning the most comprehensive set of complex reinforcement-learning tasks and verifiers released to date. The run supports heterogeneous compute—letting everyone, from consumer GPUs to hyperscale NVIDIA and AMD clusters, contribute meaningful work towards frontier-level AGI research. With TOPLOC v2 verifiable proofs, we ensure honest computation. There is no waitlist, no permission required, and no cap on how much compute you can bring. Just spin up your GPUs and start helping us advance towards open-source superintelligence. You can now contribute to by clicking “Contribute Compute” on the SYNTHETIC-2 dashboard. → https://xmrrwallet.com/cmx.plnkd.in/g9CirX8D

Prime Intellect Meetup July 18th Berlin Join us https://xmrrwallet.com/cmx.plu.ma/m86mfwof

Chip into the bigger conversation

Introducing PCCL, the Prime Collective Communications Library — a low-level communication library built for decentralized training across the globe. After six months of dedicated research led by Michael Keiblinger, we are releasing PCCL to accelerate research on distributed low-communication optimization algorithms and further close the gap to centralized training. PCCL has been designed from the ground up for fault tolerance to cleanly and quickly recover from any possible IO failure. Communication libraries like Nvidia's NCCL assume fast, stable connections—ideal for supercomputers, not the public internet. In our testing PCCL achieves up to 45 Gbit/s of bandwidth across data centers in Western Europe and 25 GBit/s training intercontinental across North America and Western Europe. If you want to see how we use PCCL, refer to our open source prime repo, which is our production-ready open source implementation of DiLoCo. Detailed Technical Report: https://xmrrwallet.com/cmx.plnkd.in/gyCKwz6M Github: https://xmrrwallet.com/cmx.plnkd.in/g8hainew Docs: https://xmrrwallet.com/cmx.plnkd.in/gy6baCZt Blog: https://xmrrwallet.com/cmx.plnkd.in/gJsuWxfW

📣 ANNOUNCED: $1M in grants for open-source AI that advances truth-seeking. Cosmos Institute x Foundation for Individual Rights and Expression (FIRE) are partnering to defend free thought and promote open inquiry in AI. AI should sharpen thought, not replace it. We’re putting our money where our mouths are: • 90-day prototypes, $1k-$10k+ per project • Rolling review with decisions in ~3 wks • Join the top AI x philosophy network • Compute powered by Prime Intellect • Meetups hosted in major talent hubs • Top projects demo at AI x Free Speech Symposium in Austin Bari Weiss announced this on stage last night at The Free Press debate, "Will the Truth Survive Artificial Intelligence?" with Fei-Fei Li, Aravind Srinivas, Jaron Lanier, and Nicholas Carr, and we published an op-ed earlier today. "The First Amendment constrains governments, but the principles of free speech must also be translated into code. We're challenging builders to do exactly that and prioritize freedom over control." Greg Lukianoff, FIRE "AI already guides one-fifth of our waking hours. The builders of these systems now hold the future of free thought and expression in their hands. This program gives them the capital, compute, and community to seize that opportunity." Brendan McCord, Cosmos Institute Let's build for freedom. Links below.

Releasing INTELLECT-2: We’re open-sourcing the first 32B parameter model trained via globally distributed reinforcement learning. Unlike traditional centralized training efforts, INTELLECT-2 trains a reasoning language model using fully asynchronous RL across a dynamic, heterogeneous swarm of permissionless compute contributors. Links Available in Comments Below 🔗 Detailed Technical Report 🔗 INTELLECT-2 on Hugging Face 🔗 Chat Interface to try it out 🔗 Prime-RL: our async RL framework