Prime Intellect logo

Prime Intellect

Prime Intellect is an active AI infrastructure company building an open stack for training, deploying, evaluating, and continuously improving agentic models.
San Francisco, United States

INTELLECT-3


Project Links


Description

An open-source 106B-parameter MoE model with 12B active parameters, post-trained from GLM-4.5-Air-Base using Prime Intellect’s prime-rl framework, verifiers environments, and large-scale RL infrastructure. The model weights, training framework, environments, and technical report are publicly available.

Category: Uncategorized

SYNTHETIC-2


Project Links


Description

Prime Intellect launched SYNTHETIC-2 as an open-source reasoning dataset and associated planetary-scale, pipeline-parallel distributed inference run. The dataset includes prompts, model completions, rewards, verified DeepSeek-R1-0528 traces for supervised fine-tuning, and traces from smaller models for reinforcement-learning difficulty estimation. It is distributed through Hugging Face, with full, SFT-verified, SFT-unverified, and RL subsets.

Category: Data

Verifiers


Project Links


Description

verifiers is Prime Intellect's framework for defining tasks, running agents and harnesses, scoring them on set tasks, and using the results for evaluations and reinforcement learning. Its current v1 architecture separates tasksets, harnesses, runtimes, and training-ready traces, and integrates with Prime Intellect's Environments Hub and prime-rl.

Category: Developer Tooling

Prime-Rl


Project Links


Description

prime-rl is Prime Intellect’s modular framework for asynchronous RL post-training. It coordinates inference, rollout orchestration, and training; supports SFT, RL, evaluations, multimodal models, verifiers environments, and multi-node deployments scaling from single GPUs to 1000+ GPUs.

Category: Developer Tooling

Prime Agent


Project Links


Description

Prime Agent is a self-improving agent harness built around Recursive Language Models and a persistent IPython kernel. It supports recursive subagents, durable sessions, a multi-process local runtime, scheduling, MCP integrations, and a public CLI. Its documentation states that it began as a hard fork of pi-mono but is now developed and distributed independently as Prime Agent.

Category: AI Agent Framework

Prime Flash MoE


Project Links


Description

Prime Flash MoE provides fused bf16 and MXFP8 MoE kernels for NVIDIA Blackwell GPUs. It fuses gate/up projections, SwiGLU activation, down projection, and top-k reduction, with fused and split execution pipelines. The repository is installable and includes API, testing, and benchmarking instructions; the kernels are integrated into Prime Intellect’s prime-rl framework.

Category: Uncategorized

Prime Sandboxes


Project Links


Description

Prime Sandboxes provides isolated, disposable Docker workspaces managed through the Prime CLI and Python SDK. They support custom container images, CPU resources, secrets, environment variables, network controls, file transfer, port exposure, SSH access, labels, timeouts, and automated lifecycle management. Prime Intellect also describes the infrastructure as an execution layer for high-concurrency reinforcement-learning rollouts and agentic workloads.

Category: Developer Tooling

Prime Inference


Project Links


Description

Prime Inference is Prime Intellect's hosted model-inference service. It supports API and Prime CLI access, model listing and selection, token-based billing, streaming, standard OpenAI parameters, team accounts, and deployment of trained LoRA adapters through an OpenAI-compatible endpoint.

Category: Developer Tooling

Prime Intellect Compute


Project Links


Description

Prime Intellect Compute aggregates GPU capacity from multiple providers and lets users provision on-demand GPU instances, multi-node clusters, persistent storage, and related orchestration and monitoring through the Prime Intellect platform, CLI, and API.

Category: Compute Network

Prime Intellect Lab


Project Links


Description

Lab combines Prime Intellect's environments and verifiers ecosystem with hosted evaluations and training workflows. Its official tooling supports setting up Lab workspaces, discovering environments, running evaluations, configuring Hosted Training, and launching training runs through the Prime dashboard and CLI.

Category: AI Agent Platform

Prime Intellect CLI & SDKs


Project Links


Description

Prime Intellect CLI & SDKs is an open-source Python developer-tooling project distributed primarily as the prime package. It provides CLI commands and SDK clients for Prime Lab, hosted reinforcement-learning training, RL environments and evaluations, GPU pods and availability, remote code-execution sandboxes, inference, teams, storage, and tunnels. The repository is maintained by the PrimeIntellect-ai GitHub organization and is actively released under the MIT License.

Category: Developer Tooling

Environments Hub


Project Links


Description

Environments Hub is Prime Intellect’s community-powered platform for aggregating, showcasing, creating, sharing, and evaluating reinforcement-learning environments. It also functions as a Python package registry for environment modules distributed as wheels and integrates with Prime Intellect’s RL training and evaluation tooling.

Category: AI Agent Platform

Lab


Project Links


Description

Lab provides an end-to-end model-improvement loop centered on environments, including hosted evaluations, reinforcement-learning training, rollout inspection, adapter deployment, and inference.

Category: AI Agent Platform