Ollama
The simplest way to get an open-weight model answering on your own hardware, and its local API has become a default target that most agent frameworks already speak. The cloud tier is the practical escape hatch when a model will not fit in your RAM. What you do not get is fleet-scale serving: this is a single-node runtime, not an inference platform with autoscaling and usage metering.
Pros
- One command pulls and runs a model, with no environment wrangling
- Local HTTP API that most agent frameworks already support
- Works fully offline once weights are downloaded
- Cloud tier covers models too large for local hardware, in US, EU and Singapore regions
- Runtime is open source under a permissive licence
Cons
- Single-node runtime with no autoscaling or multi-tenant serving
- Max tier signups were paused for capacity when checked
- Local speed is capped by your own GPU or RAM