Getting started

Models and access

Available models

River provides optimized training and inference implementations for the following open-weight models. Each implementation is tested for correctness; River handles model placement and the distributed execution behind API calls.

deepseek-ai/DeepSeek-V4-Flash-0731
nvidia/GLM-5.2-NVFP4
nvidia/GLM-5.2-NVFP4-262K
nvidia/Kimi-K2.6-NVFP4
nvidia/Kimi-K2.6-NVFP4-262K
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
Qwen/Qwen3.5-9B
Qwen/Qwen3.5-122B-A10B-FP8
Qwen/Qwen3.5-397B-A17B-FP8
Qwen/Qwen3.6-35B-A3B-FP8
Qwen/Qwen3.8-27B-FP8
XiaomiMiMo/MiMo-V2.6-Flash-RL
XiaomiMiMo/MiMo-V2.6-Pro-RL
zai-org/GLM-5.3-Flash

Access is granted per account, so not every model above is enabled for every API key.

MiMo V2.6 Flash RL

XiaomiMiMo/MiMo-V2.6-Flash-RL supports text-only sampling and LoRA training. Training supports LoRA ranks up to 32 and sequences up to 262,144 tokens. Sampling supports a 262,144-token context (prompt plus output), with at most 32,768 output tokens per request. Set train_unembed=False when creating the adapter:

model = session.create_model(
    base_model="XiaomiMiMo/MiMo-V2.6-Flash-RL",
    lora=river.LoraConfig(rank=32, train_unembed=False),
)

MiMo V2.6 Pro RL

XiaomiMiMo/MiMo-V2.6-Pro-RL supports text-only sampling and LoRA training. Training supports LoRA ranks up to 32 and sequences up to 262,144 tokens. Sampling supports a 262,144-token context (prompt plus output), with at most 32,768 output tokens per request. Set train_unembed=False when creating the adapter:

model = session.create_model(
    base_model="XiaomiMiMo/MiMo-V2.6-Pro-RL",
    lora=river.LoraConfig(rank=8, train_unembed=False),
)

For current sampling and training rates, open Billing → Pricing in the River Console.

Check what your key can use

A minimal end-to-end check: connect, confirm the server is healthy, and print the base models your key can use.

import os
import river_client as river

client = river.Client(api_key=os.environ["RIVER_API_KEY"])

print("healthy:", client.health_check())
for name in client.get_capabilities():
    print(name)

Example output:

healthy: True
Qwen/Qwen3.6-35B-A3B-FP8
Qwen/Qwen3.5-397B-A17B-FP8
nvidia/Kimi-K2.6-NVFP4
nvidia/GLM-5.2-NVFP4

get_capabilities() is the authoritative source: it returns the live list of model names your key can pass as base_model in later calls. Always read it at runtime rather than hard-coding — the catalog changes over time, and it is scoped to your account, so it can return fewer models than the catalog above.

If you'd like access to more models, reach out on Discord or email support@river.ai.

Cloud and on-premises

The River API is available in two deployment models:

	River Cloud	On-premises
GPU infrastructure	Hosted by River	Your existing GPU cluster
Training and inference	River manages the distributed execution	The same River stack runs on your infrastructure
Data and weights	Stored in River Cloud; trained weights can be downloaded	Training data, model weights, and inference run within your environment
Getting started	Create an API key in the Console	Work with River to deploy and configure your cluster

The examples in this guide use River Cloud. On-premises deployments use the same Python client with your cluster's endpoint and credentials. Available models and capacity depend on the deployment. Contact api@river.ai to discuss an on-premises installation.

For an on-premises endpoint supplied by your administrator:

client = river.Client(
    api_key=os.environ["RIVER_API_KEY"],
    endpoint="your-api-hostname",
)