> ## Documentation Index
> Fetch the complete documentation index at: https://docs.pinecone.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Fireworks AI with Nexus BYOC

> Learn how Fireworks AI open models fit into a Nexus BYOC deployment and why you'd use them for generation and curation.

> [!NOTE]
>
> Nexus BYOC is available only on [Enterprise plans](https://www.pinecone.io/pricing/).

[Fireworks AI](https://fireworks.ai/) is an inference and training platform for open models, including Kimi, DeepSeek, GLM, Qwen, MiniMax, and gpt-oss. In a Pinecone Nexus Bring Your Own Cloud (BYOC) deployment, you can serve Nexus's chat and embedding models from Fireworks.

## Fireworks AI features

* **Serverless inference:** Call models on shared capacity and pay per token.
* **Dedicated deployments:** Reserve capacity for a model when you need predictable throughput.
* **Fine-tuning:** Tune a model on your own data and serve the tuned model from the same platform.

## Nexus integration

Nexus uses models at several points, which [Nexus model guidance](https://docs.pinecone.io/guides/nexus/models.md) describes. In BYOC, the inference proxy routes every model call according to your model catalog, `inference-proxy-models.toml`. Each catalog entry names an upstream model, the endpoint that serves it, and the secret that holds its API key. Fireworks exposes an OpenAI-compatible endpoint, so a Fireworks model is a catalog entry with `api_style = "openai"` pointed at that endpoint.

Once Fireworks models are in the catalog, they behave like any other model:

* **Work-area defaults:** The catalog's `phase_defaults` assigns a default model to each work area (Query, Curation, and System). A call that names no model runs on its work area's default.
* **Per-query selection:** A query can name any enabled model by its catalog ID, and the model selector on a context's **Query** tab lists every model enabled for Query.
* **Embeddings:** You can optionally route the default embedding model to Fireworks. The setup guide keeps reranking on a Pinecone-hosted model.

## Benefits

* **Lower curation cost:** Curation runs a model across every source, which makes it the most cost-sensitive step in Nexus. A fast open model as the Curation default keeps that cost down as your corpus grows.
* **Open models for every work area:** You can run Query, Curation, and System on open models, and compare them against closed models on your own questions.
* **Fine-tuned models:** You can fine-tune a model on Fireworks and serve it to Nexus by adding it to your catalog.

## Key concepts

| Concept | Fireworks AI | Nexus BYOC |
| - | - | - |
| Model | A model path, such as `accounts/fireworks/models/kimi-k3`. | A catalog entry under `[llm_models.<id>]` or `[embedding_models.<id>]`, referenced by its catalog ID. |
| Endpoint | The OpenAI-compatible endpoint, `https://api.fireworks.ai/inference/v1`. | The `base_url` on the catalog entry. |
| API key | The key that authenticates each request. | The `nexus-provider-keys.<ref>` stack secret that the entry's `api_key_ref` names. |
| Generation | Serves the chat model that composes each answer. | The Query default, or the model a query names. |
| Curation | Serves the chat model that reads sources and builds artifacts. | The Curation default. |
| Embeddings | Serves embedding models, such as Qwen3 Embedding 8B. | The default embedding model, which is Pinecone-hosted unless you change it. |

To set up Fireworks models, see [Use Fireworks AI models in Nexus BYOC](https://docs.pinecone.io/guides/nexus/byoc/fireworks/use-models.md). For supported models and catalog fields, see the [Fireworks AI reference](https://docs.pinecone.io/guides/nexus/byoc/fireworks/reference.md).