Prime Intellect launches INTELLECT-2, a 32B reasoning model trained across the globe for ~$100K

Summary

Prime Intellect co-founder and CEO Vincent is building what he describes as an interplanetary compute cluster — a network that aggregates idle GPUs across data centers, gaming rigs, and consumer devices worldwide, and routes AI workloads across them without requiring trust in any individual node.

The company's headline this week is INTELLECT-2, a 32B reasoning model trained across globally distributed hardware for roughly $100,000 in compute. The model is built on top of Qwen and shows meaningful gains in math and coding. Vincent frames it explicitly as a proof of concept — the goal was to validate four things at once: that untrusted compute contributors can be verified at scale, that distributed reinforcement learning works with low communication overhead, that fault tolerance holds when nodes drop in and out, and that the approach can improve frontier open-source models cheaply and repeatably.

Why RL makes this tractable

The shift from pre-training to reinforcement learning post-training is what makes distributed training genuinely viable at scale. Pre-training requires tight, frequent gradient synchronization across all nodes. RL rollout generation is largely decoupled — nodes can generate samples independently and sync less often. That structural property is what let Prime Intellect run INTELLECT-2 across H100s and A100s scattered across the globe rather than inside a single data center.

The next run will go further: 3090s and consumer gaming cards will be able to contribute, with tasks routed to nodes based on difficulty. Nodes can join for a few hours when idle and drop out without breaking the run.

Verification overhead

The verification problem is where Prime Intellect's approach diverges from earlier crypto-native distributed compute networks. Cryptographic verification mechanisms carry 1x to 10x overhead, which makes them uncompetitive with centralized clusters. Prime Intellect's verification runs at roughly 1% overhead, which Vincent says is what makes the economics work for AI researchers rather than just as a blockchain use case.

Talent and positioning

Will Brown, a well-regarded distributed RL researcher, recently joined Prime Intellect. Vincent says open-source AI is attracting researchers who want to publish and build in the open — something closed labs can't offer. He argues Prime Intellect is one of the few Western organizations genuinely closing the gap with Qwen and DeepSeek, and describes Meta's Llama 4 launch as disappointing by that standard.

Hardware landscape

Vincent expects centralized mega-clusters like OpenAI's Stargate and distributed mid-tier capacity to grow simultaneously. Nvidia, he argues, has an incentive to fragment the buyer market rather than concentrate it — selling chips to CoreWeave, Lambda, and others rather than ceding leverage to AWS and the hyperscalers building their own silicon. Beyond data centers, major automakers and phone manufacturers have approached Prime Intellect about monetizing idle compute in vehicles and handsets, though Vincent says that use case is not yet operational.

DeepSeek R2

Vincent expects DeepSeek R2 to close further on frontier closed models but probably fall short of Gemini and OpenAI's top models. He views the broader open-source convergence trend as durable. Prime Intellect kicked off its largest synthetic data generation run the day DeepSeek dropped, using the resulting reasoning data for INTELLECT-2. The ambition is continuous, week-over-week improvement on open-source models rather than waiting months between major releases.