# Prime Intellect

> Prime Intellect is an AI infrastructure company that builds open-source tools and hosted services for training open models with reinforcement learning.

Source: https://metavert.io/prime-intellect  
Published: 2026-10-07  
Updated: 2026-10-07

**Prime Intellect** is an AI infrastructure company that builds open-source software and hosted services for training and improving open-weight models with [reinforcement learning](https://metavert.io/reinforcement-learning). It describes its product as "the open superintelligence stack": GPU compute, a training framework, a library and public hub for [RL environments](https://metavert.io/rl-environments), sandboxes, inference, and — since August 2026 — an open-source agent harness. This profile draws on the company's own site, blog and GitHub repositories; performance and business figures are the company's and are labelled as such.

### From Distributed Training to an RL Platform

The company's early public work was on training models across geographically dispersed hardware. Its blog documents OpenDiLoCo, an open-source framework for low-communication distributed training (July 2024), followed by INTELLECT-1, presented as a 10B-parameter model trained across globally distributed compute (launched October 2024, released November 2024), and INTELLECT-2, a 32B-parameter model trained with globally distributed reinforcement learning (April–May 2025).

From mid-2025 the emphasis shifted from [distributed training](https://metavert.io/distributed-training) as such to the tooling around RL. The Environments Hub launched in August 2025 with the argument that high-quality environments had been locked inside closed labs. INTELLECT-3 (November 2025) was a 106B-parameter mixture-of-experts model trained with supervised fine-tuning and RL on top of GLM 4.5 Air; the company says it used 512 NVIDIA H200 GPUs over two months and claims state-of-the-art results for its size, a vendor-reported claim. Lab, the hosted training platform, was introduced in February 2026 and opened in May 2026.

On funding, the company announced $15M in February 2025 and a $130M Series A led by Radical Ventures in July 2026, with NVIDIA Ventures, Intel Capital and Dell Technologies Capital participating, for a stated total of more than $150M. The Series A announcement claims more than 6,000 customers and over $100M in annualised revenue; these are company statements and have not been independently audited.

### The Open-Source Stack

| Component | What it is | Licence (per GitHub, October 2026) |
| --- | --- | --- |
| verifiers | Library for building RL environments and evals as modular components | MIT |
| prime-rl | Asynchronous RL training framework: an orchestrator, a vLLM-based inference server and a trainer | Apache 2.0 |
| Environments Hub | Public registry for creating, sharing and discovering environments | Hosted service |
| Prime Agent | Agent harness built on the recursive language model abstraction | MIT |

**verifiers** packages a task, its tooling and its scoring into one unit usable for both training and evaluation; the repository credits Will Brown as its original author. **prime-rl** is the trainer that consumes those environments. Its README describes fully asynchronous training — generation and optimisation decoupled so that slow agentic rollouts do not stall the trainer — with inference served by [vLLM](https://metavert.io/vllm), support for supervised fine-tuning, RL and evals, and both dense and mixture-of-experts architectures. The README's scale claims (1,000+ GPUs, models beyond a trillion parameters) are the project's own.

The **Environments Hub** listed more than 2,500 community environments on the company's homepage as of October 2026. Community-contributed environments vary in quality, and a count says nothing about how many have verifiers robust enough to train against without inviting [reward hacking](https://metavert.io/reward-hacking).

The commercial layer sits on top of this. **Lab** combines hosted training, hosted evaluations, adapter deployment, inference and sandboxes, and is priced per token rather than per cluster-hour; at its May 2026 opening the company said it supported 14 models from 1B to 70B parameters and that beta users had run more than 10,000 training jobs. The pattern — open libraries, paid managed infrastructure — is common among [open-source AI](https://metavert.io/open-source-ai) companies.

### Prime Agent

Prime Agent, announced on 5 August 2026 and described in an arXiv paper later that month (Karten et al.), is a coding and long-task [agent harness](https://metavert.io/agent-harness) built around two ideas. The first is the [recursive language model](https://metavert.io/recursive-language-models): the agent's only tool is a persistent IPython session in which context is held as variables and sub-agents are invoked as function calls. The second, which the company calls a Continual Harness, lets the agent create, read, update and delete its own prompts, skills, memory and sub-agent definitions based on its past trajectories — a form of [agentic memory](https://metavert.io/agentic-memory) that extends to the harness itself.

The headline result is vendor-reported: on ARC-AGI-3, the company states a best single-attempt score that it says is level with the human-expert baseline using Opus 5 as the underlying model — the paper describes this as raising the best single-attempt score from 30% — and it reports competitive or better results on a set of long-context benchmarks with several underlying models. These numbers come from the company's blog post and its own paper. As of October 2026 no independent replication was found, and because the harness runs on third-party frontier models, the results measure the scaffold and the model together rather than either alone.

### Position and Open Questions

Prime Intellect occupies the layer between raw GPU rental and closed fine-tuning APIs: it supplies the pieces needed to run [reinforcement fine-tuning](https://metavert.io/reinforcement-fine-tuning) on open models without assembling them from scratch. Alternatives exist at each layer — Hugging Face TRL and Unsloth for single-GPU RL, other open RL frameworks for larger runs — and the company's distinguishing bet is the shared environment registry. Whether community environments reach the quality that frontier labs build privately, and how the open-source components relate to the paid platform over time, are the questions a neutral observer would watch.

## Related Topics

- [RL Environments](https://metavert.io/rl-environments) — What the verifiers library and Environments Hub provide
- [Reinforcement Fine-Tuning](https://metavert.io/reinforcement-fine-tuning) — The workflow the stack is built for
- [Recursive Language Models](https://metavert.io/recursive-language-models) — The abstraction Prime Agent is built on
- [Agent Harness](https://metavert.io/agent-harness) — The category Prime Agent belongs to
- [GRPO](https://metavert.io/grpo) — The most widely used algorithm for open-model RL
- [Open-Weight Models](https://metavert.io/open-weight-models) — The models this infrastructure trains
- [vLLM](https://metavert.io/vllm) — The inference engine inside prime-rl
- [Distributed Training](https://metavert.io/distributed-training) — The company's original research focus
- [Hugging Face](https://metavert.io/hugging-face) — The adjacent open-model ecosystem

## Further Reading

- [Prime Intellect: The Open Superintelligence Stack](https://www.primeintellect.ai/) — Prime Intellect, accessed October 2026
- [Prime Agent: A Self-Improving RLM Agent](https://www.primeintellect.ai/blog/prime-agent) — Prime Intellect, August 2026
- [Prime Agent: A Self-Improving RLM Harness](https://arxiv.org/abs/2608.23552) — Karten et al., arXiv, August 2026
- [prime-rl repository](https://github.com/PrimeIntellect-ai/prime-rl) — Prime Intellect, GitHub, accessed October 2026
- [verifiers repository](https://github.com/PrimeIntellect-ai/verifiers) — Prime Intellect, GitHub, accessed October 2026
- [prime-agent repository](https://github.com/PrimeIntellect-ai/prime-agent) — Prime Intellect, GitHub, accessed October 2026
- [Environments Hub: A Community Hub To Scale RL To Open AGI](https://www.primeintellect.ai/blog/environments) — Prime Intellect, August 2025
- [INTELLECT-3: A 100B+ MoE trained with large-scale RL](https://www.primeintellect.ai/blog/intellect-3) — Prime Intellect, November 2025
- [Releasing Lab: the training platform for self-improving agents](https://www.primeintellect.ai/blog/lab-is-open) — Prime Intellect, May 2026
- [$130M Series A to Build the Open Superintelligence Stack](https://www.primeintellect.ai/blog/series-a) — Prime Intellect, July 2026