River

River provides an API for fine-tuning open-source language models on custom data using LoRA and reinforcement learning. The core workflow is a Python client that accepts your dataset, targets a base model in the vendor’s supported range (the docs describe 35B to 1T parameter models), and returns weights you own. There is no canvas to drag, no hosted chat product — this is a training API, and the output is a model artifact shaped by your data.

The differentiating claim is ownership. Most inference APIs let you call a model someone else controls. River’s framing is that the fine-tuned weights belong to you — not cached in a shared system, not subject to the provider’s update cycle overwriting your customizations. Combined with reinforcement learning support, the vendor positions this as a path to models that adapt to specific tasks rather than staying frozen at a single checkpoint.

The fit is narrow and deliberate. Developers fine-tuning for cost reduction — replacing a large general-purpose model with a smaller task-specific one — are the stated target. The tool is not agentic, does not provide a deployment runtime, and has no free tier. Teams evaluating it for a regulated environment will note that there is no self-hosted option: your data moves through River’s infrastructure, and the vendor does not describe on-premise alternatives on the public page. That is the condition under which teams route to alternatives that support private deployment.

Integration is API-first. The vendor describes a Python client as the primary interface, with pay-per-token billing, meaning training cost scales with the data volume and model size you bring — not a flat subscription. Sign-up is required before API access, and no sandbox or trial tier is described in the available documentation.