Start here for the decision-making essentials: what Prime Intellect does, who it is for, how it is accessed, and the first-party sources behind this profile.
LicenseEnvironment Hub providing access to and contribution of 2,500+ open-source RL environments
Best suited to
Source-backed fit
Teams training and deploying custom frontier and agentic models AI labs and researchers needing scalable RL environments Buyers of on-demand and reserved GPU clusters Browser and computer-use agent developers Builders of self-improving coding agents
AI infrastructure · Tool
Prime Intellect
Decision-ready
Prime Intellect is an AI infrastructure platform offering an integrated compute, training, inference, and sandbox stack. It provides globally operated GPU compute, open-source models including INTELLECT-3 (106B MoE), 2,500+ RL environments, distributed RL frameworks, and agent development tools like Prime Agent.
Short answers to the questions buyers and builders commonly ask about Prime Intellect. Each answer cites the shared ledger below, where every source is listed once.
01What does Prime Intellect say it can do?
Open Superintelligence Stack providing an integrated compute, training, inference, and san · Prime CLI for turning any task into an RL environment with init, develop, eval, and push w · Dedicated or serverless inference for custom models with native LoRA support, plus pay-per · Framework for asynchronous reinforcement learning (RL) at scale, with secure sandboxed cod
Train, deploy, and continuously improve your own models on an integrated compute, training, inference, and sandbox stack.
Training large-scale models optimized for agentic workflows, and turning production traces · Training browser and computer use agents against live web interactions
Train large-scale models optimized for agentic workflows.
04What integrations does Prime Intellect document?
Browserbase partnership providing scalable browser infrastructure for training agents · Uses NIXL (NVIDIA Dynamo) for RDMA-based weight transfer · Uses vLLM as inference engine for RL sampling · Provides verifiers v0.3.1 with allow/denylist support for sandbox egress control
Browserbase provides scalable browser infrastructure that has already powered work at labs, including Microsoft's Fara-7B and Google's Gemini 2.5 CUA.
05How can Prime Intellect be deployed or accessed?
Cloud subscriptions, Colocation plans, and hardware purchases · Each autonomous agent run launches on a GPU node with 8xH200s in a bwrap + network namespa
In providing our Services to subscribers of our Cloud and Colocation plans, or purchasers of our hardware ("Clients")
06What security or data-control information does Prime Intellect publish?
Vulnerability disclosure via security@primeintellect.ai with PGP encryption; reports ackno · Technical, administrative, and physical safeguards protect personal information against lo · Prime Sandboxes — high-throughput, isolated code execution layer bypassing the Kubernetes
Send detailed findings ONLY to: security@primeintellect.ai
This profile connects the jobs Prime Intellect is described as handling with its delivery model, access options and the subjects used to match it to related products in this directory.
Common use cases
Training large-scale models optimized for agentic workflows
Turning any task into an RL environment via the Prime CLI
Deploying custom models via dedicated or serverless inference with LoRA support
Performing asynchronous reinforcement learning at scale
Self-improving coding workflows with Prime Agent
Training browser and computer-use agents against live web interactions
Each fact points to a recorded source, making it easy to distinguish verified product information from claims that need checking.
Official website
HTTP 200 verified twice
First-party description
Prime Intellect - The Open Superintelligence Stack
Source-supported facts
Prime Intellect is branded as 'The Open Superintelligence Stack' — an integrated compute, · Operated by PrimeIntellect, Inc., headquartered in San Francisco, USA, with a New York Cit · Cloud subscriptions, Colocation plans, and hardware purchases are offered, with the GPU Cl
Capability
Open Superintelligence Stack providing an integrated compute, training, inference, and sandbox stack to train, deploy, and continuously improve custom models
Capability
Prime CLI for turning any task into an RL environment with init, develop, eval, and push workflows
Capability
Dedicated or serverless inference for custom models with native LoRA support, plus pay-per-token LoRA serving
Capability
Framework for asynchronous reinforcement learning (RL) at scale, with secure sandboxed code execution
Use case
Training large-scale models optimized for agentic workflows, and turning production traces into next training runs via custom adapters
View 32 more verified facts
Pricing
H200 on-demand GPU: $1.99/HR (additional GPU SKUs listed include B300 $4.99/HR, H100 $2.43/HR, B300 SXM6 x512 cluster $5.00/HR/GPU, and spot from $0.94/HR)
Platform
Globally operated compute spanning a single GPU to largest clusters: on-demand access to 1-256 GPUs, reserved clusters from 50+ providers, and a spot market for idle GPUs
Security
Vulnerability disclosure via security@primeintellect.ai with PGP encryption; reports acknowledged within 2 business days and critical issues resolved within 24-48 hours
Open source
Practical capabilities
What it helps with
8 documented areas
A concise view of the jobs, capabilities and integrations described in the recorded product sources.
Use case
Training large-scale models optimized for agentic workflows
Use case
Turning any task into an RL environment via the Prime CLI
Use case
Deploying custom models via dedicated or serverless inference with LoRA support
Use case
Performing asynchronous reinforcement learning at scale
Use case
Self-improving coding workflows with Prime Agent
Use case
Training browser and computer-use agents against live web interactions
Capability
Open Superintelligence Stack providing an integrated compute, training, inference, and san
Capability
Prime CLI for turning any task into an RL environment with init, develop, eval, and push w
Availability
Where it runs and where to get it
Source checked
Documented product formats, platforms and official distribution destinations. Availability can vary by region and plan.
Cost / license
H200 on-demand GPU: $1.99/HR (additional GPU SKUs listed include B300 $4.99/HR, H100 $2.43Environment Hub providing access to and contribution of 2,500+ open-source RL environmentsPrime Agent is fully open-source and installed via curl -fsSL https://app.primeintellect.aModel weights, training framework prime-rl, verifiers library, and all training/evaluationEnvironment Hub with 2,500+ open-source RL environmentsPrime Intellect open-sourced INTELLECT-3 weights, training framework prime-rl, datasets, REnvironment Hub hosting 2,500+ open-source RL environments for researchers and developers Yes — Prime Agent, INTELLECT-3, prime-rl, and verifiers are open-sourcedBuilding verifiers to allow the open ecosystem to train and evaluate models safelyEnvironment Hub hosting 2,500+ open-source RL environments for researchers and developers · Yes — Prime Agent, INTELLECT-3, prime-rl, and verifiers are open-sourced · Building verifiers to allow the open ecosystem to train and evaluate models safely
Application types
AI infrastructure stack branded as the Open Superintelligence Stack
Platforms
Globally operated compute spanning a single GPU to largest clusters: on-demand access to 1GPU Cloud accessed via the online platform at primeintellect.ai/app/signinTrained on 512 NVIDIA H200 GPUs across 64 interconnected nodes; deployment uses NVIDIA Bla
DeploymentCloud subscriptions, Colocation plans, and hardware purchases · Each autonomous agent run launches on a GPU node with 8xH200s in a bwrap + network namespa
LicenseEnvironment Hub hosting 2,500+ open-source RL environments for researchers and developers · Yes — Prime Agent, INTELLECT-3, prime-rl, and verifiers are open-sourced · Building verifiers to allow the open ecosystem to train and evaluate models safely
Model supportINTELLECT-3: 100B+ parameter Mixture-of-Experts (MoE) model trained with large-scale RL on · INTELLECT-2: the first 32B parameter model trained through globally distributed reinforcem · Example fine-tuning run used Qwen3-VL-8B-Instruct on the WebVoyager benchmark · INTELLECT-3 — a 106B parameter Mixture-of-Experts model · Supports GLM-5.2 (1.6 TB bf16 policy) for RL training with DPEP=32 vLLM inference
Data controlVulnerability disclosure via security@primeintellect.ai with PGP encryption; reports ackno · Technical, administrative, and physical safeguards protect personal information against lo · Prime Sandboxes — high-throughput, isolated code execution layer bypassing the Kubernetes
Learning curveIntermediate
Primary use casesTraining large-scale models optimized for agentic workflows, Turning any task into an RL environment via the Prime CLI, Deploying custom models via dedicated or serverless inference with LoRA support, Performing asynchronous reinforcement learning at scale, Self-improving coding workflows with Prime Agent, Training browser and computer-use agents against live web interactions, Training large-scale models optimized for agentic workflows, and turning production traces, Training browser and computer use agents against live web interactions
What to verify before adopting
Tested models have varying knowledge cutoffs that limit their access to certain papers
Evolution and major updates
Prime Intellect timeline
A concise history of software releases and material product changes. Events appear only when a dated source supports what changed.
1 dated update
Latest first · exact dates
Showing the newest updates and meaningful milestones. Open an entry for its summary and source.
Release note
Prime Intellect publishes Recursive Language Models research blog post
Open details
Prime Intellect published research on Recursive Language Models (RLM), a scaffolding approach in which an LLM manages its own context via a persistent Python REPL and parallel sub-LLM calls. The implementation is integrated into their verifiers environments an
Facts, answers, structured details, milestones and primary resource links cite this shared ledger. Each external page appears once; release tags from the same GitHub project are grouped under one release history.
1primeintellect.ai 12 facts · 3 answers · 1 snapshot source · Official site
These tools share a workflow, capability or audience with Prime Intellect. They may complement it rather than replace it, and are not presented as integrations or endorsements.
Environment Hub hosting 2,500+ open-source RL environments for researchers and developers to access and contribute
Model
INTELLECT-3: 100B+ parameter Mixture-of-Experts (MoE) model trained with large-scale RL on Prime Intellect's RL stack
Model
INTELLECT-2: the first 32B parameter model trained through globally distributed reinforcement learning, released May 11, 2025
Training data
SYNTHETIC-2 open dataset of four million verified reasoning traces spanning math, coding, puzzles, and instruction-following tasks, released on Huggingface for SFT and RL training
Company
PrimeIntellect, Inc.
Capability
Train, deploy, and continuously improve your own models on an integrated compute, training, inference, and sandbox stack
Application type
AI infrastructure stack branded as the Open Superintelligence Stack
Platform
GPU Cloud accessed via the online platform at primeintellect.ai/app/signin
Deployment
Cloud subscriptions, Colocation plans, and hardware purchases
Use case
Training browser and computer use agents against live web interactions
Integration
Browserbase partnership providing scalable browser infrastructure for training agents
Model
Example fine-tuning run used Qwen3-VL-8B-Instruct on the WebVoyager benchmark
Security
Technical, administrative, and physical safeguards protect personal information against loss, unauthorized access, destruction, misuse, modification, and improper disclosure
Headquarters
San Francisco, USA (primary office, with a New York City research role)
Program
Open research focus on Recursive Language Models (RLM) for long-context agent scaffolding
Team size
30 open roles advertised across research, engineering, compute, and growth
Capability
Integrated compute, training, inference, and sandbox stack for training, deploying, and continuously improving models
Open source
Yes — Prime Agent, INTELLECT-3, prime-rl, and verifiers are open-sourced
Model
INTELLECT-3 — a 106B parameter Mixture-of-Experts model
Model family
INTELLECT
Parameter count
106B parameters
Architecture
Mixture-of-Experts (MoE), trained with SFT and RL on top of the GLM-4.5-Air base model
Training data
GLM-4.5-Air base model; RL environments span Math, Code, Science, Logic, Deep Research, and Software Engineering
Benchmark
95.5% on ARC-AGI-3 with Prime Agent + Opus 5, surpassing the reported human expert baseline
Capability
Prime Agent — open-source self-improving coding harness built on the Recursive Language Model (RLM) and Continual Harness, with programmatic sub-agent and tool calling via a persistent IPython kernel
Platform
Trained on 512 NVIDIA H200 GPUs across 64 interconnected nodes; deployment uses NVIDIA Blackwell / Blackwell Ultra, NVIDIA Vera Rubin NVL72, NVIDIA Vera CPUs, and NVIDIA Dynamo inference
Security
Prime Sandboxes — high-throughput, isolated code execution layer bypassing the Kubernetes control plane for agentic RL rollouts
Capability
Prime Agent provides a persistent IPython kernel for AI agents to build research workflows
Capability
Measures how well frontier models can conduct autonomous AI research
AI infrastructure · ToolArthur AIArthur AI provides a unified control plane for trustworthy enterprise AI, delivering agent discovery, observability, guardrails, and governance across the full lifecycle. It supports LLM, tabular, CV, and NLP models with bias detection, cost optimization, and SOC 2 Type II compliance.Also mapped to AI infrastructure
AI infrastructure · ToolBraintrustBraintrust is an enterprise-grade AI observability and evaluation platform for tracing production LLM applications, running experiments on datasets, comparing prompts and models, and catching regressions via online scoring and quality gates.Also mapped to AI infrastructure
AI infrastructure · ToolCerebriumCerebrium is a serverless AI infrastructure platform for real-time, high-performance applications. It provides global, region-aware GPU deployment, sub-second cold starts, instant autoscaling, and hardened gVisor workload isolation, with support for voice agents, video models, LLMs, multimodal pipelines, and large-scale batch jobs.Also mapped to AI infrastructure
AI infrastructure · ToolChromaChroma is open-source Apache 2.0 search infrastructure for AI, offering serverless, scalable vector, full-text, regex, and metadata search on object storage. Available as Chroma Cloud with SOC 2 Type II, CMEK, BYOC, and multi-cloud replication.Also mapped to AI infrastructure
