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

# Fireworks

> Every Fireworks model Verlon supports — open-weight Kimi, GLM, Qwen, DeepSeek, MiniMax and more on Fireworks serverless — with quickstart and SDK compatibility.

Fireworks serves open-weight models — Kimi, GLM, Qwen, DeepSeek, MiniMax, gpt-oss, Nemotron and more — serverless, priced per token. Verlon routes to them in whichever SDK you already use.

Model ids are Fireworks' own fully qualified ids, `accounts/fireworks/models/<model>` (and `accounts/fireworks/routers/<router>` for Fireworks' routers), exactly as Fireworks' API takes them.

## Chat models

| Model ID | Type | Context |
| - | - | - |
| `accounts/fireworks/models/deepseek-v4p1-flash` | chat | 1049K |
| `accounts/fireworks/models/deepseek-v4p1-flash-1017` | chat | 1049K |
| `accounts/fireworks/models/ember-1` | chat | 1049K |
| `accounts/fireworks/models/glm-5p3` | chat | 1049K |
| `accounts/fireworks/models/glm-5p3-flash` | chat | 1049K |
| `accounts/fireworks/models/kimi-k3` | chat | 1049K |
| `accounts/fireworks/models/kimi-k3p1-pro` | chat | 1049K |
| `accounts/fireworks/models/minimax-m3` | chat | 512K |
| `accounts/fireworks/models/muse-spark-open` | chat | 1049K |
| `accounts/fireworks/models/qwen3p8-2p4t-a95b` | chat | 262K |
| `accounts/fireworks/models/qwen3p8-max` | chat | 262K |
| `accounts/fireworks/routers/glm-5p3-fast` | chat | 1049K |
| `accounts/fireworks/routers/kimi-k3-fast` | chat | 1049K |

> [!NOTE]
>
> This table is auto-generated from the Verlon model registry on every sync —
> see the [live registry](https://docs.verlon.ai/provider-compatibility/models.md#live-registry) for
> pricing, capabilities, and deprecation dates.

## Quickstart

```typescript theme={null}
import OpenAI from 'openai';

const openai = new OpenAI({
  baseURL: 'https://api.verlon.ai/v1',
  apiKey: process.env.VERLON_API_KEY,
});

const response = await openai.chat.completions.create({
  model: 'accounts/fireworks/models/kimi-k3',
  messages: [{ role: 'user', content: 'Hello!' }],
});

console.log(response.choices[0].message.content);
```

Fireworks serves only OpenAI Chat Completions. Calls through `/v1/responses`, `/v1/messages` or the Gemini route are translated to it and back — see [Any SDK, any model](https://docs.verlon.ai/provider-compatibility/translation.md). On the Gemini route, URL-encode the id: `/v1beta/models/accounts%2Ffireworks%2Fmodels%2Fkimi-k3:generateContent`.

## What's supported

> [accordion] Streaming
>
> Full support for `stream: true` in every supported SDK.

> [accordion] Tool / function calling
>
> Supported on the models that take tools, including a forced tool choice. The
> registry's `capabilities` object lists each model's support.

> [accordion] Structured output
>
> `response_format` with `json_object` or a `json_schema` on the models that
> support it.

> [accordion] Vision
>
> Image input on the vision models (the Kimi line, among others).

> [accordion] Stop sequences
>
> Several Fireworks models reason before they answer, and some write that
> reasoning into the answer itself, where a stop sequence would cut it before
> the answer starts. On those models Verlon sends the request without `stop`
> and cuts the answer at the sequence itself (`x-verlon-emulated: stop`).

> [accordion] Embeddings
>
> Fireworks' embedding models are served on `/v1/embeddings`.

> [accordion] Reranking
>
> Fireworks' Qwen3 reranker (`accounts/fireworks/models/qwen3-reranker-8b`) is
> served on `/v1/rerank` in Fireworks' own shape: send `query`, `documents`
> (strings) and optionally `top_n`, `return_documents` and `task`, and the
> documents come back most relevant first under `data`, each with its `index`
> and `relevance_score`. A client calling Fireworks' `/inference/v1/rerank`
> swaps the base URL. A rerank bills the prompt tokens Fireworks reports, at
> its published \$0.20 per million.

> [accordion] Usage + cost tracking
>
> Every response includes token usage plus Verlon's `cost` field.

### Not available

* **PDF input.** Fireworks' chat models take no documents.
* **Models without a published price.** A Fireworks model with no published serving price can't be billed on Verlon's keys and is refused there; it still serves on [your own Fireworks key](https://docs.verlon.ai/platform/byok.md).

## Related

* [Any SDK, any model](https://docs.verlon.ai/provider-compatibility/translation.md)
* [DeepSeek](https://docs.verlon.ai/provider-compatibility/deepseek.md) — DeepSeek's own API, beside its open weights here.
* [Owned models](https://docs.verlon.ai/platform/owned-models.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.