Beam vs Prime Intellect (2026)

Code on Beam reaches other systems through what the customer writes. Secrets hold credentials, S3 buckets mount into apps and custom registries supply images. Task callbacks POST results to the customer's server, and sandbox ports expose services. Beam has no catalog of connectors, OAuth actions or tool gateway of its own, and the tools are the customer's code.

Sourcedocs.beam.cloud/llms.txtread 2026-09-22

Agents built in Prime Intellect environments reach outside systems through the tools their harness gives them. Environments declare tools, sandboxes and stateful tool environments. Hosted evaluations can be granted temporary sandbox, instance or tunnel permissions for environments that use tools, and Prime Tunnel exposes local services to them. Prime Intellect names no native connectors, OAuth actions or credential scoping for business systems, so tools are code the customer writes into the environment.

Sourcedocs.primeintellect.ai/hosted-training/what-is-labread 2026-09-22

An agent framework ships in the SDK (Bot, BotLocation, BotContext) and is built on a Petri net model. Locations hold typed state, and transitions consume inputs from locations and write outputs to others. Transitions can be driven by a model or by plain code, run concurrently in isolated containers and keep state synchronized between them. Task queues add retries and timeouts, and map fans work out. Deterministic steps mix with agent steps, and branching and retries are built in. Beam says it is launching the framework and does not say how mature it is beyond the example it gives.

Sourcedocs.beam.cloud/v2/agents/introductionread 2026-09-22

Lab runs agent steps in a loop the customer defines. An environment's harness sets the model's tools, sandboxes, context management and multi turn interaction. For each Hosted Training run, a dedicated Orchestrator manages environment logic, schedules rollout requests and coordinates training. Multi node clusters can be orchestrated with Slurm. The harness and rollout orchestration serve training and evaluation and are written in code. Prime Intellect names no workflow runtime that sequences, branches, retries and routes a deployed agent's production work.

Sourcedocs.primeintellect.ai/hosted-training/what-is-labread 2026-09-22

The schedule decorator runs a deployed job on a schedule, and web endpoints and task queues boot containers from zero when requests or tasks arrive. One app can send events to another through signals, for example to reload a model when retraining finishes. Task callbacks also notify the customer's server when a task finishes.

Sourcedocs.beam.cloud/v2/function/scheduled-jobread 2026-09-22

Work on Prime Intellect starts when a person or the customer's own code calls it. Hosted evaluations and training runs launch from the dashboard or CLI, and sandboxes and instances start through the API or SDK. Prime Intellect names no schedule, event subscription or other way to start an agent's work without a caller. Cluster alerting notifies people about hardware health, not agents.

Sourcedocs.primeintellect.ai/tutorials-environments/hosted-evaluationsread 2026-09-22

Any retrieval system on Beam is one the customer builds. Beam maintains no retrieval structure of its own over the customer's content for an agent to query. Volumes, durable disks, Maps and mounted S3 buckets store data but return no search results, and nothing Beam provides indexes, refreshes or permissions the customer's sources.

Sourcedocs.beam.cloud/llms.txtread 2026-09-22

An environment's dataset is a set of tasks with optional answers for training and scoring. In the search agent guide, the customer builds its own document search inside an environment. Prime Intellect names no retrieval structure it maintains over the customer's content.

Sourcedocs.primeintellect.ai/hosted-training/what-is-labread 2026-09-22

State carries across runs. Sandbox snapshots capture filesystem and memory as an immutable artifact to fork from or restore, and distributed volumes and durable disks on each node persist data. Distributed Queues and Maps share it across tasks, and the agent framework synchronizes state between transitions, so an agent's code can read that state and carry on. Snapshot state is the machine's own state, which a customer can restore or delete but not review or edit. Beam names no memory layer for agents.

Sourcedocs.beam.cloud/v2/sandbox/snapshotsread 2026-09-22

What persists on Prime Intellect is compute state. Paused instances keep a workspace directory, and persistent disks hold datasets and checkpoints. Sandbox snapshots that save, restore and fork a sandbox mid run are marked coming soon and have not shipped. Prime Intellect names no memory an agent reads across runs.

Sourcedocs.primeintellect.ai/sandboxes/overviewread 2026-09-22

For work running on Beam, including agents built with its framework, Beam names no approval step, consent checkpoint or escalation to a person. The platform isolates workloads with containers per customer, network separation and scoped secrets.

Sourcedocs.beam.cloud/v2/security/securityread 2026-09-22

Prime Intellect names no approval step, consent checkpoint or escalation to a person for the agents that run in its environments or sandboxes. The developer configures network allow and deny lists per sandbox, which restrict where an agent's code can reach.

Sourcedocs.primeintellect.ai/sandboxes/overviewread 2026-09-22

A SOC 2 Type II audit by Advantage Partners covers Beam, with the report under NDA, and Beam offers HIPAA BAAs for Enterprise customers and a GDPR DPA. Data is encrypted in transit and at rest, and containers and networks are isolated per customer. The customer scopes, rotates and deletes its encrypted secrets, and workspace API tokens can be created, rotated and revoked. All members of a workspace currently have the same permissions, and single sign on is not yet available.

Sourcedocs.beam.cloud/v2/security/securityread 2026-09-22

Sandboxes are microVMs isolated at the hardware level, with network allow and deny lists for egress. Secrets are injected encrypted and never appear in logs or API responses. API keys carry scoped permissions, such as Inference, and team members hold Admin or Member roles. Prime Intellect names no SSO, SCIM or audit log and publishes no SOC 2 or ISO attestation.

Sourcedocs.primeintellect.ai/sandboxes/overviewread 2026-09-22

Application logs are kept 30 days on Developer and Team plans and one year on Growth, and log export is included. Task status can be queried by ID, with callbacks on completion. Prompts, tool calls, retrieved knowledge and outputs can be inspected step by step only as far as the customer's code logs them. Beam names no audit log of workspace actions separate from the runtime logs.

Sourcebeam.cloud/pricingread 2026-09-22

Each run leaves a record. Hosted evaluations store every run with its results in Prime Evals, stream logs to the CLI and dashboard, and can be shared. Hosted Training reports run metrics. Reserved clusters come with Grafana dashboards and alerting on GPU, node and network health. Prime Intellect names no audit log, step by step trace retention or SIEM export.

Sourcedocs.primeintellect.ai/tutorials-environments/hosted-evaluationsread 2026-09-22

There are three ways to run Beam. The managed cloud runs on AWS and Google Cloud. With bring your own cloud, Beam runs the control plane and compute in the customer's AWS or GCP account for a flat management fee, and the compute bills to the customer's cloud. The pricing plans add Azure as a bring your own cloud option.

Beta9, the engine that powers Beam, is open source under AGPL 3.0 and can be self hosted on a local machine or Amazon EKS. Compute pools can also run Beam workloads on the customer's own machines. The beta9 repository is actively maintained. Beam names no region selection for the managed cloud.

Sourcedocs.beam.cloud/v2/self-hosting/overviewread 2026-09-22

Customers choose where their Prime Intellect compute runs. Before provisioning, GPU availability can be filtered by region (for example United States, Canada or Europe West), by data center and by provider cloud, across more than fifty providers. The training and evaluation stack (prime-rl and verifiers) is open source and runs on the customer's own hardware. Hosted Training and Hosted Evaluations run on infrastructure Prime manages, and no region is published for them.

Sourcedocs.primeintellect.ai/api-reference/check-gpu-availabilityread 2026-09-22

Ready made starting points ship as end to end examples for model serving (Llama, DeepSeek, Qwen and Hugging Face models, plus an OpenAI compatible vLLM server), for media (ComfyUI, Whisper and text to speech), and for fine tuning and web scraping, plus a research assistant agent built on the agent framework. Developers build from these examples, and Beam offers no browsable catalog of agents a customer can adopt as they are.

Sourcedocs.beam.cloud/v2/examples/overviewread 2026-09-22

The Environments Hub is a community registry of environments, ready to run, for training and evaluation. They span math, coding, games, search and tool use, and multimodal tasks, and the Hub also holds validated benchmark implementations. Lab guides give training recipes for code generation, tool use and search agents. Environments are tasks and scoring for research. The Hub carries no production workflows or agents built for a specific role.

Sourcedocs.primeintellect.ai/tutorials-environments/environmentsread 2026-09-22

Customers choose the model, and Beam runs any model on the GPU the customer picks. Examples serve Hugging Face models such as Llama 3.1, DeepSeek R1 and Qwen through vLLM, including an OpenAI compatible vLLM server, or through SGLang, and others fine tune Gemma and Llama. Customers can bring their own deployments and providers.

Sourcedocs.beam.cloud/llms.txtread 2026-09-22

The customer chooses the model throughout. Prime Inference is an OpenAI-compatible API that routes requests to a range of model providers. Evaluations run against any model on it or any OpenAI-compatible endpoint the customer brings. Hosted Training lets the customer pick among open weights models (Qwen, Llama, NVIDIA Nemotron, gpt-oss and others), each with its own token price. Trained LoRA adapters deploy behind an OpenAI-compatible endpoint.

Sourcedocs.primeintellect.ai/inference/overviewread 2026-09-22

Outside callers reach Beam through Python and TypeScript SDKs, a CLI, and a REST API with published Swagger specifications for the gateway and pods, covering tasks, pods and sandboxes. Web endpoints expose the customer's code as HTTP APIs, the open source Beta9 engine can be extended, and a docs MCP server serves Beam's documentation to coding tools.

Sourcedocs.beam.cloud/llms.txtread 2026-09-22

Outside code reaches Prime Intellect through a REST API (instances, availability, disks, clusters, sandboxes and chat completions) with scoped API keys, the prime CLI and a Python SDK for sandbox lifecycles. With the open source verifiers and prime-rl libraries, customers build and publish their own environments to the Environments Hub. The API, CLI and SDK are stable. Prime Intellect publishes no MCP server.

Sourcedocs.primeintellect.ai/llms.txtread 2026-09-22

Code, experiments and fine tuning jobs run on Beam with retries and logs. Beam names no fixtures, scoring or quality gates for a customer's agent workflows, and no way to compare agent runs.

Sourcedocs.beam.cloud/llms.txtread 2026-09-22

Evaluating the customer's agent is a core product. An environment packages a dataset of tasks, the harness the model runs in (tools, sandboxes, multi turn interaction) and a rubric of reward functions and metrics that scores it.

Hosted Evaluations run that environment against Prime Inference or any OpenAI-compatible endpoint, from the dashboard or CLI, and store each run with its logs and results in Prime Evals. The Environments Hub adds validated benchmarks such as AIME, MATH-500 and Humanity's Last Exam. Teams can test an agent on these datasets before production and score it the same way on every run.

Sourcedocs.primeintellect.ai/tutorials-environments/hosted-evaluationsread 2026-09-22

Sandboxes run any code with GPUs, networking and storage and can expose ports for web services. Beam names no browser or desktop for an agent to operate and no API for screenshots, clicks or typing.

Sourcedocs.beam.cloud/v2/sandbox/overviewread 2026-09-22

Browser use exists only inside training environments. The verifiers library's BrowserEnv lets an agent navigate and act on web pages, either in a DOM mode through Stagehand or in a computer use mode based on vision. The browser itself runs on Browserbase, a separate service with its own account and billing, sold beside the platform and not part of it. Prime Sandboxes are VMs for running code. Prime Intellect hosts no browser or desktop of its own for an agent to operate on web interfaces or desktop environments.

Sourcedocs.primeintellect.ai/guides/browser-environmentsread 2026-09-22