canonical: https://jentic.com/apis/climate.api.nvidia.com/nvidia-earth-2

# Climate Api Nvidia Earth-2 Inference Server

Jentic publishes the only available OpenAPI specification for Earth-2 Inference Server, keeping it validated and agent-ready. NVIDIA's Earth-2 Inference Server hosts a portfolio of NVIDIA NIM and partner models behind a single API: weather forecasting, vision-language models, image and video generation, biology models including ESMFold and DiffDock, genomics pipelines, retrieval embeddings and reranking, chat completions, and the cuOpt optimisation engine. The 19 endpoints together let developers route a request to the right model and poll long-running jobs via `/status/{requestId}.`

## For AI agents

Run NVIDIA-hosted models for weather forecasting, vision-language reasoning, image and video generation, biology, genomics, retrieval, and route optimisation through one bearer-authenticated API.

## Scope

Does not handle GPU provisioning, model fine-tuning, or training jobs - use for hosted inference against NVIDIA-curated models on the Earth-2 server only.

## Capabilities

- Generate weather forecasts via `/nvidia/ai-weather-forecasting` and ensemble weather models
- Run chat completions against hosted language models via `/chat/completions`
- Generate images and video via `/genai/stabilityai/sdxl-turbo`, `/genai/stabilityai/stable-diffusion-xl`, and `/genai/stabilityai/stable-video-diffusion`
- Reason over images via vision-language models such as `/vlm/microsoft/kosmos-2` and `/vlm/nvidia/neva-22b`
- Predict protein structures via `/biology/nvidia/esmfold` and dock molecules via `/biology/mit/diffdock`
- Run genomics pipelines via `/genomics/nvidia/deepvariant` and `/genomics/nvidia/fq2bam`
- Compute retrieval embeddings and rerank results via `/retrieval/nvidia/embeddings` and `/retrieval/nvidia/reranking`, and solve routing problems via `/nvidia/cuopt`

## Use cases

### Weather and Climate Forecasting

Generate medium-range and ensemble weather forecasts for a region by calling `/nvidia/ai-weather-forecasting` and the related forecast endpoints, which run NVIDIA's AI weather models on Earth-2 infrastructure. Agents can request a forecast, poll `/status/{requestId}` until the job is complete, and retrieve the predicted fields.

Example prompt: Submit a weather forecast request via `/nvidia/ai-weather-forecasting` for a specified bounding box and lead time, then poll `/status/{requestId}` until the predicted fields are available.

### Drug Discovery and Protein Modelling

Combine `/biology/nvidia/esmfold` to predict a protein structure from sequence with `/biology/mit/diffdock` to dock candidate ligands and `/biology/nvidia/molmim/generate` to suggest novel candidates. The biology endpoints together support a structure-based screening workflow without local GPU infrastructure.

Example prompt: Predict a protein structure with `/biology/nvidia/esmfold`, then dock a SMILES ligand against the predicted PDB via `/biology/mit/diffdock` and rank the resulting poses.

### Multimodal Generative Workflows

Run image, video, and language workflows from a single endpoint family - generate images with SDXL Turbo or Stable Diffusion XL, produce short clips with Stable Video Diffusion, and caption or reason over images with the VLM endpoints. The `/chat/completions` endpoint handles the language side so a single agent can orchestrate end-to-end multimodal pipelines.

Example prompt: Generate an image with `/genai/stabilityai/stable-diffusion-xl`, then caption it via `/vlm/microsoft/kosmos-2` and feed the caption into `/chat/completions` for downstream summarisation.

### Retrieval-Augmented Generation Backbone

Build the embedding and reranking backbone for a RAG application using `/retrieval/nvidia/embeddings` to vectorise documents and `/retrieval/nvidia/reranking` to rerank candidate passages against a user query. Combined with `/chat/completions`, this is a complete NVIDIA-hosted RAG stack.

Example prompt: Embed a corpus via `/retrieval/nvidia/embeddings`, retrieve top-K candidates from a vector store, rerank them via `/retrieval/nvidia/reranking`, then call `/chat/completions` to answer the user query.

### AI Agent Multimodel Orchestration via Jentic

An AI agent that needs weather, vision, biology, or routing capability can search Jentic, find the matching Earth-2 endpoint, and execute it without holding the raw NVIDIA API key. Jentic securely stores the bearer token and the agent only sees scoped execution.

Example prompt: Use the Jentic search query 'predict a protein structure' to find `/biology/nvidia/esmfold` and execute with the input amino acid sequence.

## Key endpoints

| Method | Path | Description |
| --- | --- | --- |
| POST | `/chat/completions` | Run chat completions against hosted language models |
| POST | `/nvidia/ai-weather-forecasting` | Generate AI-driven weather forecast |
| POST | `/biology/nvidia/esmfold` | Predict a protein structure from sequence |
| POST | `/biology/mit/diffdock` | Dock a ligand against a protein structure |
| POST | `/genai/stabilityai/stable-diffusion-xl` | Generate an image from a text prompt |
| POST | `/retrieval/nvidia/embeddings` | Generate text embeddings |
| POST | `/retrieval/nvidia/reranking` | Rerank candidate passages against a query |
| GET | `/status/{requestId}` | Poll status of a long-running inference job |

## Key resources

- **Forecast** — AI weather forecasting and ensemble forecast endpoints
- **Chat** — Chat completions against hosted language models
- **Multimodal** — Image and video generation plus vision-language reasoning
- **Biology** — ESMFold protein structure prediction, DiffDock ligand docking, MolMIM molecule generation
- **Genomics** — DeepVariant variant calling and fq2bam genomics pipelines
- **Retrieval** — Embeddings and reranking endpoints for RAG
- **cuOpt** — Routing and scheduling optimisation
- **Status** — Poll the status of long-running inference requests

## Why Jentic

- **Setup:** Wiring the Earth-2 Inference Server by hand means managing its bearer token against climate.api.nvidia.com/v1 and polling `/status/{requestId}` for long-running inference yourself. Through Jentic you install once, import the Earth-2 Inference Server from the API Directory, store the token once, and your agent calls it.
- **Permission scoping:** Earth-2 takes the model inputs in the request body of each inference endpoint, so scope the agent to the operations it needs, such as weather forecasting or protein structure prediction. You choose which operations are in the allowed set, so other hosted models are not callable unless you add them.
- **Credential handling:** Your NVIDIA Earth-2 bearer token is stored once, encrypted, by your own Jentic One instance and injected at execution time. It never enters the agent's prompt, logs, or context.
- **Discovery method:** Agents search Jentic by intent such as 'predict a protein structure' or 'forecast the weather', and Jentic returns the matching Earth-2 operation with its input schema so the agent calls the right endpoint without browsing the reference docs.

## Related APIs

- **OpenAI** — Hosted language, embedding, image, and audio models
- **Replicate** — Hosted inference for open-source models with a marketplace surface
- **Hugging Face** — Hosted models, datasets, and inference endpoints

## FAQ

### Why is there no official OpenAPI spec for Earth-2 Inference Server?

NVIDIA does not publish a single consolidated OpenAPI specification for the Earth-2 Inference Server portfolio. Jentic generates and maintains this spec so that AI agents and developers can call Earth-2 Inference Server via structured tooling. It is validated against the live API and kept up to date. Get started with Jentic One, the self-hosted execution layer.

### What authentication does the Earth-2 Inference Server use?

The API uses a Bearer token in the Authorization header obtained from NVIDIA. Through Jentic the bearer token is stored encrypted in the vault and the agent receives a scoped execution context only.

### Can I run protein structure prediction with this API?

Yes. POST a sequence to `/biology/nvidia/esmfold` to receive the predicted structure, and chain into `/biology/mit/diffdock` for ligand docking on the predicted PDB.

### How do long-running inference jobs work?

Submit the inference request to the relevant model endpoint, then poll GET `/status/{requestId}` with the returned request ID until the result is ready. This pattern applies to weather, biology, and other heavyweight models.

### How do I generate an image through Jentic?

Search Jentic for 'generate an image', load the schema for `/genai/stabilityai/stable-diffusion-xl`, and execute with the prompt. The Python SDK uses await client.search, await client.load, await client.execute.

### Which retrieval endpoints power a RAG stack?

Use `/retrieval/nvidia/embeddings` to vectorise text and `/retrieval/nvidia/reranking` to rerank candidate passages against a query. Combined with `/chat/completions` you have a complete NVIDIA-hosted RAG flow.

### Can I limit what my agent is allowed to do with the Earth-2 Inference Server API?

Yes. Because you run Jentic One yourself, your own rules decide which Earth-2 operations and credentials the agent may use, and you put only the endpoints it needs in the allowed set. For example, you can let the agent call weather forecasting via `/nvidia/ai-weather-forecasting` or protein structure prediction via `/biology/nvidia/esmfold` while leaving the other hosted models uncallable. Any operation you do not add, such as image generation via `/genai/stabilityai/stable-diffusion-xl` or chat completions via `/chat/completions`, stays off limits until you choose to include it.