Resources /
What’s new at Prime Intellect. Releases across evaluations, training, environments, inference and sandboxes.
Training runs now have a default overview with key metrics and persistent chart pins. Pin the charts you want to highlight from the Metrics page, and keep your run tabs stable across reloads.
Run headers provide easier access to monitoring links, with clearer run-starting states and GPU counts in the training lists.
Manage API keys programmatically with public API operations to list, disable and delete keys. Existing keys also support expiry-date editing.
Use scoped permissions and expiry dates to limit credential exposure. Review the current API reference for authorization requirements before managing keys.
Attach named persistent storage to sandboxes so files can outlive an individual VM. Create, attach and manage volumes through the Prime CLI and Python SDK.
Follow the documented mount and sharing rules. Volume storage has a separate lifecycle and billing: deleting a sandbox does not mean all retained storage has been deleted.
Prime Traces now scrubs recognized credentials from uploaded traces, helping reduce accidental exposure of API keys and secrets in recorded agent interactions.
Redaction is not a guarantee that all sensitive information is removed. Keep secrets out of prompts, logs and trace content, and review private data before uploading it.
Our updated open-source training and evaluation stack introduces INTELLECT Policy Optimization (IPO) as the default RL loss, alongside training performance improvements and expanded model support.
verifiers 0.4.0 unifies the stack around v1 task, environment, harness and runtime interfaces. The legacy v0 implementation is removed: review the migration and breaking changes before upgrading existing environments.
The joint announcement was published October 19, following the earlier package releases.
Pause a dedicated training run and resume it from its saved state and volume. Release GPU compute while paused, then continue the run when capacity is available.
Pause support arrived October 15, followed by stored-state resume and the prime train resume command on October 17–18. This is training checkpoint recovery, not sandbox memory/process pause-resume. Retained storage can continue to incur charges.
Connect sandboxes over private LAN networks with rules controlling which members can communicate. Manage networks through the CLI and SDK without exposing tool servers to the public internet.
Follow-up releases add joining a network when a sandbox is created and configuring rules at network creation. Each sandbox retains its own resource limits and lifecycle.
Launch full-parameter training on Prime-managed GPUs with per-GPU-hour billing, without reserving a dedicated cluster. Select a model and capacity, submit a training configuration, and monitor the run from the CLI or dashboard.
On-demand Hosted Training is in beta and requires approved access and a funded wallet. Configure checkpoint storage before launching if you need saved model weights.
The shared hosted LoRA service no longer accepts new training runs. Use on-demand Hosted Training or a dedicated Model Factory cluster for new workloads.
Prime Agent has been rewritten in Rust, with a faster, cleaner and more reliable implementation for coding and research workflows.
The v0.10.0 release brings the new implementation to users. Read the announcement for the architecture, experiments and measured results, and the repository releases for installation and version-specific changes.
Store, view, and search training and evaluation traces with API access and native support for the verifiers v1 trace format. Use the existing run viewers or browse the Traces page.
Platform runs upload traces by default:
To stop uploads, select your workspace and turn off Upload runs to Prime Traces under Billing → Usage Limits (personal) or Team Billing → Usage Limits. Existing traces are kept.
Legacy evaluation and training sample storage retires on November 6, 2026. Data will not migrate automatically.
Select each workspace, open personal settings or team settings, select Request export, then return when ready to select Download archive before the deadline.
Export and download are free. Archives cannot be imported directly into Prime Traces.
Clef and Clef Flash are available through Prime Inference and OpenRouter for evaluation and decision workloads.
Use the Evaluations API for supported scoring and classification tasks rather than the Chat Completions API. Check the current catalog for model availability, capabilities and pricing.
Our inference offering is now publicly available. Access frontier models through the Prime CLI or an OpenAI-compatible API using your Prime API key, with serverless endpoints and reserved capacity.
The serving infrastructure behind this release has powered large-scale customer deployments, including Trinity, since January, alongside our internal RL training, evaluations, and agent workloads.
Select Prime Inference as your provider to use Prime-hosted models directly in your agent workflows.
Read the Prime Inference announcement
Save a sandbox’s filesystem as a checkpoint and restore it into another sandbox. The CLI and Python SDK support checkpoint creation, restoration and waiting for a checkpoint to become durable.
Filesystem checkpoints do not preserve running processes or memory. Checkpoint storage is billed separately from sandbox compute; delete retained checkpoints when you no longer need them.
Run supervised fine-tuning configurations through prime train on dedicated training infrastructure. Use persistent volumes for datasets and model checkpoints, and inspect progress through the CLI and run dashboard.
Hosted SFT is in beta. Request access and follow the current setup guide for configuration, dataset upload and supported models.
Keep datasets and checkpoints in named volumes and attach them to training runs with prime train --volume. Volumes can be managed independently of a run, with SSH access and file-transfer tools added in subsequent updates.
Configure checkpoint saving for the weights you want to keep, and review volume capacity and storage billing before launching. Volumes are available in beta; access and supported cluster configurations apply.
Prime Sandboxes are now generally available. Run agents in full Linux virtual machines, using the Docker images and development tools you already work with.
Use them through the CLI and SDK, or as part of your evaluation and training workflows with verifiers and prime-rl.
$ prime sandbox create python:3.11-slim$ prime sandbox run <sandbox-id> -- python3 -c "print('hello')"Each sandbox has its own Linux kernel, with native support for Docker Compose, background jobs, and system-level tools. Bring an existing container image to get started.
Run v1 environments on Prime’s hosted evaluation infrastructure. Evaluations execute in VM sandboxes, with traces available to inspect on the platform.
The Environment Hub also identifies v1 and Legacy runtimes, helping you find the right environment for your workflow.
For releases before September 14, 2026, explore our announcements on X, documentation, and repository release logs.