Modal Training
Managed training runs on Modal's fleet, configured in Python alongside the rest of a team's code; also gives instant access to GPU compute, including RDMA-enabled GPUs.
Find alternatives to Modal Training
What it does
- Model training
Features 5
- Data pre-processingTurns raw data into training data at any scale with a single function.
- Supervised fine-tuningBrings a team's own fine-tuning stack and scales it from one GPU to a cluster.
- Multi-node GPU clustersConfigures multi-GPU nodes and RDMA-connected clusters for distributed training.
- CheckpointingSaves checkpoints to Volumes to resume training, run evals or serve the best one.
- Disaggregated reinforcement learningRollouts, inference and training run as separate services with their own autoscalers.
Sources
- Pricing
-
Metered per second of compute on Modal's one rate card — Nvidia B200 $0.001736/sec, H200 SXM $0.001261/sec, H100 SXM5 $0.001097/sec, A100 80GB $0.000694/sec — with $30 a month of free compute. The pro
- Description
-
Official documentation · 5 Sep 2026
Modal's own product page for it, fetched 2026-09-05, HTTP 200.
- Constraints
-
Official documentation · 2 Oct 2026
Derived from the security facts on this page, from the same source.
- Security
-
Official documentation · 2 Oct 2026
Platform-wide: security, data-residency, region-selection, audit-logs, SCIM and RBAC docs. HIPAA BAA is Enterprise plan only and excludes Volumes v1, Images (most), Memory Snapshots and user code. SSO, SCIM and audit logs are Enterprise-plan features; SCIM is beta. ZDR unknown at platform level: only Inference endpoints are described as zero data retention. Training is from the Modal terms: without prior written consent Modal will not 'train any AI model using Customer Data' (this page adds that Modal will not access inputs, outputs or stored data). Customer Supplied Encryption Keys exist but are Alpha, so customer-managed-keys not recorded. Trust portal trust.modal.com is script-rendered.
- Sold within
-
Official documentation · 5 Sep 2026
Modal's pricing page prices compute, not products: every one of the six entries in its Products nav runs on the same account and the same meter.
- Interfaces
-
Official documentation · 8 Oct 2026
"Run pip install modal to install the modal Python package" -- a Python SDK and a CLI workflow for every product on the same account/meter; modal-notebooks additionally has a browser notebook UI.
- Intended for
-
"Where researchers can run experiments, not ops"; footer "AI infrastructure that developers love."
Also from Modal
-
Modal Infrastructure service
Serverless GPU compute: a Python decorator puts a function on an accelerator, scales it from zero to thousands of containers and stops billing when it stops running — aimed at inference, fine-tuning and batch jobs rather than reserved clusters.
-
Modal Inference Infrastructure service
Serve, scale and optimise model inference on Modal's runtime, with sub-second cold starts and autoscaling across regions.
-
Modal Sandboxes Infrastructure service
Isolated, instantly-started containers for running untrusted or agent-generated code at scale — the primitive behind AI app-generation products.
-
Modal Notebooks Developer tool
Hosted notebooks backed by Modal's GPUs, for profiling and experimenting without provisioning a machine.
-
Modal Batch Infrastructure service
Batch execution of large jobs across Modal's fleet, described as one line of code on the product page.
Something wrong here? Send a correction — quote this product id: modal-training.