Prime Intellect

Prime Intellect is a full-stack platform for training, evaluating and serving your own AI models: a multi-provider GPU marketplace, hosted reinforcement-learning post-training through its Lab product, a community Environments Hub, and inference. It is for AI engineering teams that want custom agentic models without depending on a frontier lab's closed API.

Prime Intellect is a San Francisco company, founded in January 2024 by Vincent Weißer and Johannes Hagemann, that sells the infrastructure for training and serving your own AI models rather than renting a frontier lab's API. It began in July 2024 as a compute exchange that aggregates GPU capacity from many clouds; today its homepage lists on-demand H100, H200, B200 and B300 instances, spot capacity and reserved InfiniBand clusters with Slurm or Kubernetes orchestration. On top of that sits Lab, launched from private beta on February 10, 2026 after more than 3,000 RL runs, which bundles hosted reinforcement-learning training, evaluations, sandboxes for code execution and inference (dedicated, LoRA pay-per-token and serverless OpenAI-compatible) around the Environments Hub, a community registry of RL environments opened in August 2025. The software underneath is open source: prime-rl, an Apache-2.0 asynchronous RL trainer built on FSDP2 and vLLM, and verifiers, an MIT library for environments and evals. Its model line runs from INTELLECT-1 (10B, November 2024) and INTELLECT-2 (32B, May 2025) to INTELLECT-3, a 106B mixture-of-experts released in November 2025 and post-trained from Z.ai's GLM-4.5-Air on 512 H200 GPUs across 64 nodes, a centralized cluster despite the company's decentralized-training origins. In July 2026 it raised a $130 million Series A led by Radical Ventures at a reported $1 billion valuation, disclosing an annualized revenue run rate above $100 million and customers including Ramp and Zapier. It competes with GPU clouds such as Lambda and RunPod and with managed fine-tuning platforms such as Together AI.

Category

Infrastructure & Cloud

Pricing

Usage-based, Contact for pricing

Target Market

CTOs, ML Engineers, AI Researchers, Enterprise Developers

Deployment

Multi-cloud, Open-source, API-based

Founded

2024

Headquarters

San Francisco, United States

Customers

6,000+ (company-reported, July 2026)

• ✓
  Multi-provider GPU marketplace

  Rents on-demand, spot and reserved NVIDIA GPUs sourced from many providers through one account, so teams compare prices without negotiating separate cloud contracts.

• ✓
  Lab hosted RL training

  Runs reinforcement-learning post-training, evaluations and inference as a managed service, removing the need to operate your own GPU cluster for agent training.

• ✓
  Environments Hub

  A community registry of 2,500+ RL environments built on the verifiers library, reusable both as training tasks and as benchmark evaluations.

• ✓
  prime-rl open-source trainer

  Apache-2.0 asynchronous RL framework using FSDP2 and vLLM that scales from a single GPU to 1,000+ GPUs, so workloads are portable off the platform.

• ✓
  Code-execution sandboxes

  Secure remote sandboxes execute model-generated code during training and evaluation, which coding and tool-use environments need to score model outputs.

• ✓
  Inference deployment

  Serves trained models through dedicated deployments, pay-per-token LoRA adapters or serverless OpenAI-compatible endpoints, closing the loop from training to production.

• ✓
  Multi-node clusters

  InfiniBand-connected clusters with Slurm or Kubernetes orchestration and Grafana monitoring support large distributed pre-training and RL jobs.

• •
  Train a task-specific agent

  Post-train an open-weight model with RL on your own environment to outperform general frontier models on a narrow task; Ramp reports its spreadsheet-analysis agent did so faster and cheaper.

• •
  Burst GPU capacity for experiments

  Rent a single H100 or a discounted spot instance for short fine-tuning or evaluation jobs, paying per minute of runtime instead of committing to a reserved contract.

• •
  Model evaluation before selection

  Run hosted evaluations of open models against Hub environments to compare candidates before choosing which base model to invest post-training effort in.

• •
  Large-scale distributed training

  Reserve InfiniBand multi-node clusters with Slurm or Kubernetes for multi-week pre-training or RL runs that outgrow a single eight-GPU node.

• •
  Reduce frontier API dependency

  Replace a closed-model API call with a smaller custom model served on dedicated or serverless inference, gaining control over weights, per-request cost and data handling.

GPU compute is usage-based: credits are deducted per minute while instances run, and the homepage listed H100 at $2.43/hr on-demand and $0.94/hr spot in September 2026, with rates varying by provider. Spot is up to 90% cheaper but interruptible, pods are deleted if credits run out, sample multi-node H100 clusters run about $40.80-$52.80/hr, and there are no formal SLAs. Lab training and dedicated inference list prices are not published; inference is per-token and larger commitments go through a sales call.

This page was written from 13 sources, 6 on domains other than primeintellect.ai.

1. 1.primeintellect.ai — primeintellect.aivendor
2. 2.primeintellect.ai — labvendor
3. 3.primeintellect.ai — series avendor
4. 4.primeintellect.ai — intellect 3vendor
5. 5.primeintellect.ai — environmentsvendor
6. 6.primeintellect.ai — fundraisevendor
7. 7.primeintellect.ai — computevendor
8. 8.docs.primeintellect.ai — faq
9. 9.github.com — PrimeIntellect ai
10. 10.github.com — prime rl
11. 11.techcrunch.com — prime intellect raises 130m series a to help enterprises bui
12. 12.implicator.ai — prime intellects intellect 3 open source ambition meets cent
13. 13.en.wikipedia.org — Prime Intellect