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://lnkd.in/gyCKwz6M Github: https://lnkd.in/g8hainew Docs: https://lnkd.in/gy6baCZt Blog: https://lnkd.in/gJsuWxfW

Introducing Michael Keiblinger's work on "Recurrence-Complete Frame-based Action Models" A paper on why long-horizon perception requires rethinking recurrence.

With the US falling behind on open source models, one startup has a bold idea for democratizing AI: let anyone run reinforcement learning.

We present to you the unified Prime Intellect infrastructure stack Now available on GitHub https://lnkd.in/gqj-F82b

Excited to share: I’ve uploaded my first RL environment to Prime Intellect’s Environments Hub - med-spell-count-qa. It’s a lightweight, objective task for GRPO training: count specific letters in real medical terms. Paired with the MedSpellCountQA dataset on Hugging Face, it’s a clean benchmark for string-reasoning and attention under RL. If you’re exploring GRPO or want fast, reproducible RL evals beyond math-only tasks, this is for you. 🚀

We are launching Reserved Instances. Prime Intellect’s internal research team has benefited from using our global network of cloud partners to secure GPU clusters, dynamically monetizing idle compute in our ecosystem, and mitigating risk as we scale. Now we’re providing this capability to the broader community. Prime Intellect reserved instances transform how GPUs are bought, used, and sold. Request 8–1,000+ GPU clusters and receive quotes from 50+ providers in under 24 hours. Receive immediate access to secure H100s, H200s, B200s, and NVL72 clusters. Monetize idle reserved GPUs through our spot market, reducing costs for reserved instance buyers, spot users, and providers. Additional Features - Monitoring with Grafana dashboards - Dedicated support from our research team - Orchestration with SLURM, Ray or Kubernetes - Native integrations: Environment Hub, Sandboxes, RFT, Multi-Node Training Submit a request today: https://lnkd.in/gtHtxwgw

For over a year now I've been the sole growth personnel at Prime Intellect. A lot has changed since I first started but the goal remains the same, create open source intelligence. Unlike traditional growth roles focused on enabling sales at Prime Intellect growth enables research. From our own lab's research to external labs using our ◦ GPU Cloud Marketplace ◦ Distributed Training ◦ Environment Hub Today Prime Intellect is hiring for three roles on the growth team. We are looking for individuals who understand the risks of centralized super intelligence and have deep expertise for their respective roles to drive change. If this is you or someone you know, please see the 3x jobs linked below ↓

OpenBench 0.4.0 is here! 🚀 We've collaborated with Prime Intellect, Rootly, Vercel, and many others to bring some powerful new features to OpenBench. Here's what's new: 🤝 Prime Intellect Integration We've partnered with Prime Intellect to make OpenBench evaluations directly runnable as reinforcement learning (RL) environments. You can now define your evaluation once, seamlessly push it to Prime Intellect’s Environments Hub, and immediately begin training models tailored specifically to your tasks. This transforms evaluations into actionable training loops, bridging the gap between testing and training. 🧩 GPT-OSS Scoring via OpenAI In collaboration with OpenAI, we've integrated their GPT-OSS scoring patterns into popular benchmarks like MMLU, GPQA Diamond, and MathArena (AIME, BRUMO, HMMT). Now, you can precisely replicate OpenAI’s evaluation methods across any model or provider, ensuring more accurate and reliable comparisons. 🔐 CTI-Bench: Cybersecurity Evaluations Introducing CTI-Bench (h/t Parker Johnson) We're also adding CTI-Bench, an eval for evaluating LLMs in cyber threat intelligence, to OpenBench. This is the first step we're taking towards adding OpenBench into the AI Evaluations ecosystem that David O. Sacks, Howard W. Lutnick, and the White House have been building for safe, transparent AI assessment. 🚒 Rootly GMCQ: Real-world SRE Benchmark Thanks to Rootly, we've added GMCQ, a benchmark explicitly designed to evaluate real-world Site Reliability Engineering (SRE) tasks- incident triage, log analysis, and outage mitigation. This marks the first time we've collaborated directly with benchmark creators, Laurence Liang and Sylvain Kalache, to open-source their evaluation through OpenBench. 🚀 Vercel AI Gateway Integration Our collaboration with Vercel has enabled integration with their AI Gateway, allowing evaluations directly against any model available through their router. Simplified access and seamless performance are now just a command away. We're excited to keep growing the OpenBench ecosystem, thanks to our incredible community and partners. We’re rapidly iterating and actively hiring - if you want to contribute directly, the best way to join us is by submitting pull requests from our GitHub issues list. Try OpenBench 0.4.0 today (link in comments), and don't forget to star us on GitHub if you find it valuable! ⭐