Training

Infrastructure for training

Manage rollouts, evals, and clustered training from a single file.

Get started Read the docs

Agent infrastructure that agents love.

Data pre-processing

vol:/data

s3 . hf . volume put

raw/*.json

512 containers

vol:/data

train/*.bin

Turn raw data into training data at any scale, with a single function.

Get started

Parallelize with a single call

Fan out massive preprocessing jobs across containers with `.spawn_map()`
Match compute to your pipeline

Run processing on GPUs or CPUs for disaggregated, scalable pipelines.
Built-in distributed storage

Store raw and processed data in Volumes for reuse, or BYO cloud bucket.

Supervised Fine-Tuning

vol:/data

s3 . hf . volume put

train/*.bin

Containers on 4 nodes

vol:/data

train/*.bin

Bring your fine-tuning stack and scale it from a single GPU to a cluster.

Get started

Bring your training stack

Run TRL, Unsloth, or custom PyTorch code in Modal containers.
Scale from one GPU to a cluster

Configure multi-GPU nodes and RDMA-connected clusters for distributed training.
Write checkpoints

Save checkpoints to Volumes to resume training, run evals, or serve the best one.

Reinforcement Learning

TRAINER

8 containers . 1 node

Inference

6/14

Sandboxes

2,048

Rollouts that scale fast without slowing down your workflow.

Get started

Disaggregated by default

Rollouts, inference, and training run as separate services with their own autoscalers.
Battle-tested with Open Source

Run the open RL frameworks trusted by researchers.
Native sandbox environments

Write to a Modal Volume for easy access for RL, evals, or serving.

Engineered for the speed and scale agents require

On-demand access to single and multi-node GPUs

Instant access to the compute shape you need, including RDMA-enabled GPUs, with a few lines of code.

Learn more

Managed GPU and cluster health

Spend more time experimenting and less time managing the state of the GPUs you’re working with.

Learn more

Battle-tested infra on Open Source frameworks

Tune everything, from post-training techniques to OSS frameworks and hyperparameters.

Learn more

Fine-Tuning

“By providing high-GPU-concurrency sweeps, persistent volumes, and observability tools, engineers could easily launch, monitor, and compare fine-tuning runs without worrying about cluster orchestration or storage plumbing.”

Cyrus Asgari Research Engineer, Decagon

Evals

“Modal Sandboxes let us recreate complex, production-like systems - from multi-container infrastructure workloads to GPU-enabled deep learning environments - across the millions of simulations we run each month.”

Rustem Feyzkhanov Sr. Engineering Manager — AI Platform, Snorkel AI

ML Pipelines

“Modal lets us deploy new ML models in hours rather than weeks. We use it across spam detection, recommendations, audio transcription, and video pipelines, and it’s helped us move faster with far less complexity.”

Mike Cohen Head of AI & ML Engineering, Substack