River AI

River AI is a training-and-inference API that lets engineering teams fine-tune open-weight language models with LoRA and reinforcement learning, then serve the result, billed per million tokens instead of per GPU hour. It is aimed at enterprises that want a customised model they own rather than a rented closed-source endpoint, and it removes the need to stand up and babysit a distributed training cluster.

River AI is a token-metered training and inference API for open-weight large language models, built by xAI co-founder Igor Babuschkin alongside a team drawn from Google DeepMind, OpenAI, xAI and Tesla. The company was incorporated in Nevada on 20 April 2026, came out of stealth in June 2026 from Palo Alto, California, and announced $1.1 billion across combined Series Seed and Series A rounds on 11 August 2026, led by General Catalyst and AMP PBC with strategic investment from NVIDIA and AMD Ventures and participation from Y Combinator and Temasek. The shipping product is the River API, currently marked preview v0.1: a small Python client installed with pip install river-client that exposes low-level training primitives rather than a single high-level train() wrapper. Developers run supervised fine-tuning and reinforcement learning against a catalogue of open-weight base models spanning roughly 35B to 1T parameters, including Qwen3.6 35B, Qwen3.5 397B, Kimi K2.6, GLM 5.2, DeepSeek V4 Flash and NVIDIA Nemotron 3.5 Lightning 30B, using LoRA adapters with configurable rank from 1 to 32. The documented surface covers forward-backward passes, optimiser steps, sampling, checkpoint save and resume across sessions, and five loss functions: cross_entropy, importance_sampling, ppo, cispo and dro. River runs the GPU workers behind an asynchronous submit-then-poll interface and absorbs weight transfer, sampling-training consistency and elastic compute. Billing is metered per million tokens for both training and inference rather than per GPU hour, so idle capacity is never charged, and trained adapters deploy directly to production inference on the same platform.

Category

AI Models & APIs

Pricing

Usage-based

Target Market

CTOs, VPs of Engineering, ML Engineers, AI Platform Teams, Data Scientists

Deployment

Cloud-only, API-based

Founded

2026

Headquarters

Palo Alto, California, United States

• ✓
  LoRA fine-tuning on 35B-1T open-weight models

  Adapter-based supervised fine-tuning with configurable rank 1-32 across Qwen3.6, Qwen3.5 397B, Kimi K2.6, GLM 5.2, DeepSeek V4 Flash and Nemotron 3.5 Lightning, so teams pick a base sized to the task.

• ✓
  Reinforcement learning with five loss objectives

  Policy-gradient training via importance sampling, plus cross_entropy, ppo, cispo and dro losses, letting teams optimise against custom reward functions rather than only imitating a labelled dataset.

• ✓
  Token-metered billing for training and inference

  Charges are per million tokens rather than per GPU hour, so an engineering team pays for work actually done and never for reserved capacity sitting idle between experiments.

• ✓
  Low-level Python client, not a black-box trainer

  The river-client SDK exposes forward-backward, optimiser step, sampling and checkpointing directly, so researchers keep control of the training loop while River operates the distributed backend.

• ✓
  Managed distributed training backend

  River handles fast weight transfers, sampling-training consistency and elastic compute across GPU workers behind an asynchronous submit-then-poll API, removing the cluster operations burden entirely.

• ✓
  Instant deployment of trained checkpoints

  Training and inference checkpoints are distinguished, sessions can be resumed, and a finished adapter serves production traffic on the same platform without an export or redeploy step.

• •
  Domain-specialised assistant on proprietary corpora

  Fine-tune an open-weight base on internal documentation or support transcripts so answers reflect house terminology, and keep the resulting adapter weights in-house.

• •
  RL on verifiable tasks such as maths or code

  Optimise a model against programmatic rewards using the documented GSM8K-style policy-gradient recipe, improving task accuracy beyond what supervised imitation of examples achieves.

• •
  Cost reduction versus a closed frontier API

  Replace a large general-purpose model with a smaller specialised one for a high-volume narrow workload, cutting per-token inference cost while holding task accuracy.

• •
  Agent tool-calling reliability tuning

  Train a base model on an organisation's specific tool schemas and failure traces so agents call the right internal APIs with correctly shaped arguments more consistently.

• •
  Rapid experimentation without cluster procurement

  Run and compare several reinforcement-learning configurations in a single afternoon, since no GPU reservation, scheduling or fault-recovery engineering has to be built first.

River publishes real per-token rates rather than hiding behind a sales call, and meters both training and inference on tokens instead of GPU hours, so idle capacity costs nothing. Training runs from $1.00 per million tokens on Qwen3.6 35B and Nemotron 3.5 Lightning up to $10.00 on Qwen3.5 397B, with Kimi K2.6 reported at $12.84 per million at 262k context; cached prompt tokens bill at 20% of the prompt rate and checkpoints cost $0.10 per GB per month. There is no published free tier, no committed-spend or enterprise tier, and no volume-discount schedule — the API is still preview v0.1 and access starts by signing up at console.river.ai.

This page was written from 7 sources, 4 on domains other than river.ai.

1. 1.river.ai — river.aivendor
2. 2.river.ai — apivendor
3. 3.river.ai — series seed series a fundingvendor
4. 4.docs.river.ai — docs.river.ai
5. 5.techcrunch.com — general catalyst leads 1 1b round into 2 month old river ai
6. 6.thenextweb.com — river ai 1 1 billion babuschkin general catalyst open weight
7. 7.unite.ai — river ai raises 1 1b out of stealth to rebuild the stack for