MODEL FACTORY SERVICE TERMS

1DEFINITIONS AND SCOPE

1.1 Definitions. In these Model Factory Terms:

“Base Model” means a pre-trained or previously trained model, including an open-weight model, from which a Run or other training workload starts, whether made available from the Model Cache or provided by Customer.

“Beta Features” means Hosted RL, Hosted SFT, Volumes and any other feature of Model Factory that Prime Intellect identifies in the Documentation or on the Platform as “beta,” “preview,” “early access” or by a similar designation.

“Checkpoint” means a saved copy of model weights and, where applicable, optimizer and training state, produced during or at the end of a Run or by a workload on Dedicated Capacity.

“Dedicated Capacity” means GPU capacity that is reserved for or dedicated to Customer under an Order.

“Hosted Training” means Prime Intellect’s on-demand, managed training service, including supervised fine-tuning (“Hosted SFT”) and full-parameter reinforcement learning (“Hosted RL”), through which Customer dispatches Runs to GPU capacity operated by Prime Intellect.

“Model Cache” means the Base Models that Prime Intellect caches on its training clusters and makes available for dispatch, as listed through the Prime CLI from time to time.

“Model Factory” means Hosted Training, Volumes, Prime Traces, the Model Factory Stack and Model Factory Endpoints.

“Model Factory Endpoint” means an inference endpoint that Customer deploys on Dedicated Capacity using the Model Factory Stack.

“Model Factory Stack” means Prime Intellect’s software stack for operating Dedicated Capacity for training and inference, including managed Slurm, managed Kubernetes, shared storage mounts, monitoring and alerting, and deployment of models as Model Factory Endpoints.

“Prime CLI” means Prime Intellect’s command-line interface for accessing and operating the Platform.

“prime-rl” means Prime Intellect’s open-source training framework, and “verifiers” means Prime Intellect’s open-source environment framework, each as made available under its applicable open-source license.

“Prime Traces” means Prime Intellect’s service for storing, viewing and managing evaluation, training and agent traces (“Traces”).

“Run” means a single training job dispatched by Customer through Hosted Training, from dispatch until it reaches a terminal state (completed, failed or stopped) or is paused. A resumed Run is a new Run.

“Run Configuration” means the configuration files, parameters, deployment sizing, environment selections, secret references and other instructions that Customer provides to dispatch a Run or to run a workload using the Model Factory Stack.

“Trained Model” means the model weights, Checkpoints, adapters and other model artifacts produced by a Run or by Customer’s workloads using the Model Factory Stack, including any such artifact after conversion to another format, but excluding the Base Model as it existed before training and excluding any Prime Intellect Property.

“Volume” means a named, persistent storage volume provisioned for Customer on a cluster operated by Prime Intellect.

“Wallet” means the prepaid balance associated with Customer’s Account, together with any overdraft allowance and credit grants applied to it.

1.2 Service Categories. Hosted Training and Prime Traces are Lab Services; Volumes are Compute Services; the Model Factory Stack is a Managed Service; and Model Factory Endpoints are Inference Services provided on dedicated capacity. Model Factory is a Platform Service and not a Professional Service. Any post-training, fine-tuning or other work performed for Customer by Prime Intellect Personnel requires a separate Order for Professional Services.

1.3 Order of Precedence. With respect to Model Factory, these Model Factory Terms prevail over the Agreement, and the commercial terms of an Order (including its Order Details) prevail over these Model Factory Terms. Nothing in these Model Factory Terms overrides the provisions of the Agreement governing confidentiality, indemnification or limitation of liability.

2GENERAL TERMS

2.1 Access. Some Model Factory features must be enabled by Prime Intellect for Customer’s Account, or for a team within it, before Customer may use them. Prime Intellect may condition enablement on Customer funding its Wallet, completing Prime Intellect’s sanctions, export control and know-your-customer screening, or both.

2.2 Customer Responsibilities. Customer is solely responsible for: (a) its Run Configurations, including deployment sizing, Base Model selection, hyperparameters and environment selection; (b) enabling checkpointing and attaching a Volume wherever it wishes to retain model weights, as described in Section 3.4; (c) retrieving logs and other artifacts before teardown; (d) maintaining a sufficient Wallet balance; (e) the security of the credentials, SSH keys and API keys it supplies or uses; (f) converting Checkpoints into any format it requires; and (g) stopping or pausing Runs and deleting Volumes and Traces it no longer needs, since Fees continue to accrue until it does.

2.3 Beta Features. Beta Features are provided “as is,” are excluded from any service level or service credit, and may be modified, suspended or discontinued by Prime Intellect upon at least 30 days’ notice to Customer, or such shorter notice as is reasonably practicable where required for security, legal or third-party supplier reasons. If Prime Intellect discontinues a Beta Feature without making a materially equivalent replacement available, Prime Intellect will refund any prepaid credits or prepaid Fees that Customer can no longer use for any Model Factory service as a result. A feature ceases to be a Beta Feature when Prime Intellect makes it generally available.

2.4 No Outcome Commitment. Except as expressly set out in an Order, Prime Intellect makes no commitment, and assumes no liability, with respect to the performance, capability, accuracy, benchmark result, reward, convergence or other training outcome of any Run, Trained Model or workload, or that any Run will complete. Fees for GPU-hours consumed are payable whether or not a Run completes or achieves Customer’s objectives.

2.5 Usage Limits. Model Factory is subject to the usage limits set out in the Documentation, which may include limits on the number of GPUs per Run, the number of queued Runs, the number and size of secrets supplied to a Run, Volume size and Prime Traces storage. Prime Intellect may raise any limit on Customer’s request, and may change default limits upon notice, which may be given through the Documentation.

2.6 Third-Party Integrations. Third-party services that Customer connects to Model Factory, such as experiment-tracking services and model or dataset repositories, are Customer-Provided Products. Data that Customer directs to be sent to any such service leaves Prime Intellect’s control when sent, and Prime Intellect has no responsibility for it.

2.7 Secrets. Credentials and other secrets that Customer supplies to a Run are stored as a secret specific to that Run, are made available only to that Run’s pods, and are not carried over to any resumed or subsequent Run. Customer should supply scoped, revocable credentials and is responsible for rotating them.

3HOSTED TRAINING

3.1 Dispatch. Customer dispatches Runs through the Prime CLI using a Run Configuration in the prime-rl configuration format. A Run Configuration that fails validation (for example, because of a schema error, an unavailable Base Model or a deployment exceeding Customer’s limits) is rejected before any GPU is reserved, and no Fees accrue for it.

3.2 Capacity. Unless a Run is placed on Dedicated Capacity, Hosted Training is provided on on-demand capacity, subject to availability and without any capacity reservation. A Run that cannot be placed immediately is queued and starts automatically when capacity becomes available, and Customer’s Wallet is checked again at that time. Where Customer has Dedicated Capacity of a matching GPU type, Runs will use it before on-demand capacity. A paused Run is resumed as a new Run on whichever cluster has capacity at the time, and Prime Intellect does not commit that capacity, or the same GPU type, will be available when Customer resumes.

3.3 Base Models. Runs may start from a Base Model in the Model Cache or from a Base Model that Customer stages on a Volume. Prime Intellect may add Base Models to, or remove them from, the Model Cache from time to time, provided that removal does not affect a Run already in progress. Customer may request that Prime Intellect add a Base Model to the Model Cache, but Prime Intellect is not obligated to do so.

3.4 Checkpointing and Teardown. Checkpointing is disabled by default. Unless Customer enables checkpointing in its Run Configuration and attaches a Volume to the Run, model weights produced by a Run are not retained after the Run ends. When a Run reaches a terminal state, Prime Intellect automatically tears down the Run’s compute resources, and trainer, inference and environment-server logs held on those resources are deleted. Orchestrator logs, metrics, samples and progress information for the Run remain available through the Platform in accordance with the Documentation. Prime Intellect has no liability for the loss of any model weights, logs or other data that Customer did not save to a Volume or retrieve before teardown.

3.5 Images and Framework Versions. Runs use Prime Intellect’s managed prime-rl container image by default. Customer may select another image tag that Prime Intellect publishes for use with Hosted Training. Prime Intellect may deprecate image tags upon notice and is not obligated to maintain any particular image version.

4FEES AND BILLING

4.1 Rates. Fees for Model Factory are charged at the rates set out in the applicable Order or, if none are set out, at Prime Intellect’s then-current rates displayed on the Platform (including through the Prime CLI), including per-GPU-hour rates for Hosted Training, per-GiB storage rates for Volumes, and per-GiB ingestion and storage rates for Prime Traces.

4.2 Metering. Hosted Training Fees accrue per GPU-hour for each GPU allocated to a Run, from the time the Run’s pods are ready until the Run reaches a terminal state or is paused, prorated to the second, with no minimum. Time during which a Run is queued, being scheduled or pulling container images is not billed; time during which a Run’s pods are running to download or load model weights, or during which a Run is pausing (including saving a Checkpoint), is billed. The rate displayed when a Run is dispatched applies for the duration of that Run. Volume Fees accrue on provisioned capacity from creation of the Volume until its deletion, whether or not any Run uses it. Prime Traces ingestion Fees are charged at ingestion, and storage Fees accrue hourly on retained data.

4.3 Dedicated Capacity. Unless an Order provides otherwise, Runs placed on Customer’s Dedicated Capacity, and Customer’s use of the Model Factory Stack and Model Factory Endpoints on Dedicated Capacity, incur no GPU-hour Fees in addition to the Fees for that Dedicated Capacity under the applicable Order. Volume and Prime Traces Fees apply in all cases.

4.4 Rate Changes. Prime Intellect may change its Model Factory rates upon notice, which may be given by updating the rate displayed on the Platform (including through the Prime CLI). A changed rate applies only to Runs dispatched, and storage or ingestion occurring, after the change takes effect. This Section 4.4 applies notwithstanding any provision of the Agreement requiring advance notice of Fee increases or deferring them to a renewal term, but does not apply to rates fixed in an Order, which remain fixed for the term stated in that Order.

4.5 Wallet; Payment. Runs are admitted against Customer’s spendable Wallet balance, which consists of its balance plus any overdraft allowance and credit grants, and credit grants (including prepaid credits) are applied before Customer’s balance. If Customer’s spendable balance is insufficient, Prime Intellect may decline to start a queued Run and, after notice to Customer, may pause or stop running Runs. Usage-based Model Factory Fees are paid through the Wallet or, where Prime Intellect has approved invoiced billing, invoiced monthly in arrears, notwithstanding any provision of the Agreement providing for invoicing in advance.

5VOLUMES AND PRIME TRACES

5.1 Volumes. Each Volume is provisioned in Customer’s team or personal namespace on a specific cluster, and Runs that use a Volume are placed on that cluster. Volumes may be enlarged but not reduced, up to the maximum size stated in the Documentation. A Volume persists until Customer deletes it.

5.2 Deletion of Volumes. Deletion of a Volume is permanent and removes all data on it, including the outputs of every Run that used it. Prime Intellect cannot recover a deleted Volume. Team Volumes may be deleted only by a team administrator. Customer is responsible for designating its administrators, managing access to its Volumes and maintaining any backups it requires.

5.3 Access Sessions. Customer may access a Volume through SSH sessions over the public internet using SSH keys that Customer registers. Customer is responsible for safeguarding those keys and for closing sessions in the manner described in the Documentation so that data is synchronized.

5.4 Prime Traces. Upload of Traces to Prime Traces is enabled by default and is billable. Customer may disable uploads at the workspace level or for individual Runs or clients in the manner described in the Documentation. Disabling uploads does not delete Traces already stored, which continue to accrue storage Fees until Customer deletes them.

5.5 Retrieval After Termination. Any retrieval period that the Agreement provides after expiration or termination applies to Customer Property stored on Volumes and in Prime Traces, and storage Fees continue to accrue during it. Customer may ask Prime Intellect to delete that Customer Property before the retrieval period ends.

6MODEL FACTORY ON DEDICATED CAPACITY

6.1 Scope. Where Customer uses the Model Factory Stack on Dedicated Capacity, the applicable Order governs the Dedicated Capacity itself, including its hardware, location, Fees, term and any service levels, and these Model Factory Terms govern the Model Factory Stack. Any service level in the Order applies to the Dedicated Capacity; no separate service level applies to the Model Factory Stack unless the Order expressly provides one.

6.2 Included Components. Subject to the configuration of the Dedicated Capacity, the Model Factory Stack includes: (a) Slurm, preconfigured for single-tenant use, and Kubernetes; (b) shared storage mounted on all nodes; (c) monitoring dashboards covering GPU, node, interconnect and Slurm metrics, GPU health monitoring (including XID error detection), and alerting to destinations that Customer designates; (d) the ability to deploy models as Model Factory Endpoints; and (e) the ability to run a specific version of prime-rl selected by Customer in place of Prime Intellect’s managed build.

6.3 Allocation of Responsibilities. Prime Intellect is responsible for the physical hardware and data center facilities (directly or through its suppliers); firmware, drivers and base operating system images; the Slurm and Kubernetes control planes; the monitoring infrastructure; and diagnosing and repairing or replacing failed nodes and components. Customer is responsible for its workloads, jobs, containers and software; any configuration changes it makes; any framework version it selects under Section 6.2(e), which Prime Intellect supports on a commercially reasonable efforts basis only; its users and their access; any custom alerting rules it creates; and the data it stores on Dedicated Capacity, including any backups not expressly provided for in the Order.

6.4 Administrative Access. Where an Order grants Customer administrative access, Customer will not modify, disable or remove any component that Prime Intellect manages under Section 6.3, including drivers, firmware, monitoring agents, Prime Intellect’s endpoint detection and response tool and control-plane configuration, except as agreed in writing with Prime Intellect. Any unavailability, malfunction or Data Incident to the extent caused by Customer’s use of administrative access is excluded from any service level and from Prime Intellect’s security obligations, and Prime Intellect may, upon notice to Customer, restore any such component to its baseline configuration.

6.5 Hardware Protection. Whether or not Customer has administrative access, Customer will not disable thermal protection, power management or other hardware protection features, or operate any node outside the specifications required to maintain the hardware manufacturer’s warranty.

6.6 Model Factory Endpoints. Model Factory Endpoints run on, and consume, Customer’s Dedicated Capacity. Customer is responsible for configuring access controls and API keys for each Model Factory Endpoint and for all traffic to it. Customer may make Model Factory Endpoints available to its own end users only where the applicable Order or the Agreement permits Customer to use the Services to power its own products and services.

6.7 End of Order. Upon expiration or termination of the Order for the Dedicated Capacity, Customer’s access to the Model Factory Stack on that capacity ends, subject to Section 5.5. After any applicable retrieval period, Prime Intellect will securely erase Customer Property from the storage of that Dedicated Capacity before reallocating it.

7MODELS, DATA AND OWNERSHIP

7.1 Base Models and Third-Party Materials. Base Models in the Model Cache, environments in the Environments Hub that Prime Intellect did not author, and datasets Customer obtains from third-party sources are Third-Party Materials. Customer is responsible for complying with the license terms of each Base Model it uses, including any terms that apply to derivative models, such as attribution, naming, acceptable use and user-threshold requirements. The availability of a Base Model in the Model Cache is not a representation that Customer’s intended use of that Base Model or of any Trained Model is permitted.

7.2 Customer Property. Run Configurations, datasets, Base Models provided by Customer, environments and reward functions authored by or for Customer, and other materials that Customer uploads, stages or provides in connection with Model Factory are Customer Property.

7.3 Trained Models and Run Outputs. Trained Models, and the evaluation results, metrics, rollouts, samples and Traces generated by Customer’s Runs and workloads, are Output and are Customer’s Confidential Information. As between Customer and Prime Intellect, Customer owns all Trained Models, subject to the rights of the licensors of any Base Model or other Third-Party Materials and to Prime Intellect’s rights in the Prime Intellect Property. To the extent Prime Intellect acquires any right, title or interest in a Trained Model, Prime Intellect hereby assigns it to Customer. Prime Intellect will not access Trained Models except as necessary to provide the Services, to respond to Customer’s support requests, as required by law or as Customer instructs.

7.4 Prime Intellect Property. The Platform, the Prime CLI, Prime Intellect’s hosted training and cluster infrastructure, the Model Factory Stack, Prime Intellect’s managed container images and its monitoring tooling and configurations are Prime Intellect Property. prime-rl and verifiers are made available under their open-source licenses, and nothing in these Model Factory Terms limits Customer’s rights under those licenses with respect to the open-source code; Prime Intellect’s hosted, managed and non-public versions, configurations and infrastructure remain Prime Intellect Property.

7.5 No Training on Customer Data. Prime Intellect will not use Customer Property, Trained Models or Traces to train, fine-tune, post-train, distill or evaluate any machine learning model for Prime Intellect’s own benefit or for the benefit of any third party, except (a) as necessary to provide Model Factory to Customer, (b) in the form of aggregated and de-identified usage information that does not reveal the content of any Customer Property, Trained Model or Trace, or (c) as Customer expressly authorizes in writing.

7.6 Personal Data. Customer will not include Personal Data in any dataset, environment, Run Configuration or Trace except as the Agreement permits, and, where Personal Data subject to Privacy Laws requiring a written data processing agreement is included, only where such an agreement is in effect between the parties.

7.7 Trade Controls. Trained Models and Checkpoints may be subject to Trade Control Laws. Customer is responsible for determining the classification of, and obtaining any license or other authorization required for, any Trained Model it downloads, deploys, publishes or otherwise makes available, and Customer’s trade control obligations under the Agreement apply to Trained Models.

8UPDATES

8.1 Updates. Prime Intellect may update these Model Factory Terms from time to time by posting the updated version on the Website with a new “Last updated” date. For material changes, Prime Intellect will also notify Customer by email or through the Platform at least 30 days before the changes take effect, except where a change is required by law.

8.2 Effect of Updates. Where the Agreement governs updates to Service Terms, the Agreement applies. Otherwise, an update takes effect on the later of the date stated in the notice and 30 days after notice, and Customer’s continued use of Model Factory after that date constitutes acceptance of the update; provided that, where these Model Factory Terms are incorporated into an Order under a Signed Agreement, an update that materially reduces Customer’s rights will not apply during the then-current term of that Order unless it is required by law or Customer agrees to it. No update applies to any dispute that arose before the update took effect.