Prime Intellect
Open model training platform · Infrastructure
How they use Nebius
Prime Intellect uses Nebius capacity for distributed model training and reinforcement learning. Nebius's customer story says the team trained Synthetic 2 and Intellect 2 using its infrastructure and has coordinated jobs across Nebius data centres in Europe and the US. It calls Nebius the team's go-to cloud for flexible on-demand hardware. The story describes a distributed training workflow that runs model instances across onboarded clusters for rollouts and experiments, then aggregates the rollouts for centralized training. David Schulman’s post describes training INTELLECT-3, a 100B+ parameter MoE model, on Nebius AI Cloud. Prime Intellect also uses Nebius inference for INTELLECT-3, including an OpenRouter route named in its research. The team uses Nebius Managed Soperator for automated recovery, scheduling and topology-aware job placement. Nebius reports consistent 100% GPU utilization during training runs without advance GPU bookings. The customer story also describes using SkyPilot to manage workloads across clouds, with Nebius as a testing ground. Nebius says its platform supports scaling to thousands of GPUs. The story describes BF16 as the team's standard for distributed training and FP8 for most inference tasks, including suitable inference tasks within reinforcement learning.
The team tested a four-node, 16-GPU GB200 cluster on Nebius for training and inference before moving some pre-training work to B300 systems. The GB200 test was a proof of concept. Its TorchTune-based setup required dependency updates for the Arm-based Grace CPU but no code changes. Nebius says reinforcement-learning and pre-training experiments on the GB200 cluster led to downstream releases of INTELLECT-3 and Trinity Large, with Trinity Large trained in partnership with Arcee AI. Nebius says later pre-training used B300 systems. Following the proof of concept, Prime Intellect plans to offer GB200 NVL72 compute access through its platform.
Nebius also names Prime Intellect as a post-training and reinforcement-learning partner in its September 2026 AI Builder Program announcement.
Sources
4 sourcesRecursive Language Models: the paradigm of 2026
1 Jan 2026 · Prime Intellect · Web page
Prime Intellect reports running INTELLECT-3 through OpenRouter’s nebius/fp8 route in experiments comparing standard language models with recursive language models.
Open original sourceINTELLECT-3: A 100B+ MoE trained with large-scale RL
26 Nov 2025 · Prime Intellect · Web page
Prime Intellect names Nebius and Parasail as inference providers for INTELLECT-3.
Open original sourceShares the GB200 proof of concept for large distributed workloads
13 Jan 2026 · Dschul (LinkedIn) · Web page
Prime Intellect describes a GB200 proof of concept for distributed training on Nebius.
Open original sourceCompany context
primeintellect.aiPrime Intellect provides computing infrastructure and software for AI startups, research teams and enterprises to train, deploy and improve their own models and agents. Its Lab platform combines hosted training, reinforcement learning environments, evaluations and secure code-execution sandboxes. It also offers GPU compute, reserved clusters and inference services through a shared API. Its open-source tools include the Verifiers environment library and Prime-RL training framework. Customers include Ramp, which uses Lab to train an agent for spreadsheet search.
Recent updates
13 updatesPrime Intellect serves Moondream 3.1 on OpenRouter
Moondream 3.1 became available on OpenRouter with inference provided by Prime Intellect. The vision-language model has 9 billion total parameters, 2 billion active parameters and a 32,000-token context window.
OpenRouterPrime Intellect and NVIDIA develop enterprise tasks for Nemotron training
Prime Intellect and NVIDIA built 2,500 verified enterprise training tasks for Nemotron 3.5 Super VL. They reported that reinforcement learning increased its score on the public AutomationBench-AA split from 34.0% to 57.7% after approximately 11 hours of training.
Prime IntellectPrime Intellect acquires Moondream
GeekWire reported that Prime Intellect acquired vision AI startup Moondream for undisclosed terms. Co-founders Jay Allen and Vik Korrapati are joining to lead inference work and expand multimodal models. Prime Intellect’s own announcement appeared on October 13.
GeekWire