Baseten (US) is a model serving platform for production. Users deploy their own models or fine-tuned checkpoints and receive inference endpoints with automatic GPU scaling. The platform handles packaging, runtime and capacity, so models can be published without assembling the infrastructure yourself.
What it does
- Deploys custom models and fine-tuned checkpoints in production
- Serves inference over an API with GPU scaling
- Runs open models published in the library
- Chains inference steps into workflows
How it works
- The model is packaged with its runtime and dependencies
- The platform publishes an endpoint and adjusts replicas under load
- Dedicated deployments isolate capacity for predictable workloads
- Models can be fine-tuned and versioned inside the environment
Models and infrastructure
- Open models from the library plus customer-owned models
- Dedicated GPUs, with an exclusive deployment option
- Billed per GPU minute and per token on API models
- Separate environments for development and production
Public pricing
- Per-GPU-minute billing for dedicated deployments
- Public examples: L4 at US$ 0.01414/min, H100 at US$ 0.10833/min and B200 at US$ 0.16633/min
- API models billed per token, with a published price per model
- Startup plan with usage-based pricing and no platform fee
Strengths
- Production focus with scaling and dedicated deployments
- Accepts custom models, not just a ready catalog
- Reproducible deployment workflow
Points of attention
- Requires knowledge of model packaging and runtime
- Costs depend on GPU type and runtime
- Features vary by plan