Add the fireworks AI provider (RooCodeInc#6652)

* add fireworks provider

* add tests

* Update packages/types/src/providers/fireworks.ts

Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com>

* fix typo

* another typo

---------

Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com>

Author: ershang-fireworks

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -25,6 +25,7 @@ body:
         - AWS Bedrock
         - Chutes AI
         - DeepSeek
+        - Fireworks AI
         - Glama
         - Google Gemini
         - Google Vertex AI

packages/types/src/global-settings.ts

@@ -193,6 +193,7 @@ export const SECRET_STATE_KEYS = [
 	"codebaseIndexMistralApiKey",
 	"huggingFaceApiKey",
 	"sambaNovaApiKey",
+	"fireworksApiKey",
 ] as const satisfies readonly (keyof ProviderSettings)[]
 export type SecretState = Pick<ProviderSettings, (typeof SECRET_STATE_KEYS)[number]>
 

packages/types/src/provider-settings.ts

@@ -37,6 +37,7 @@ export const providerNames = [
 	"cerebras",
 	"sambanova",
 	"zai",
+	"fireworks",
 ] as const
 
 export const providerNamesSchema = z.enum(providerNames)
@@ -263,6 +264,10 @@ const zaiSchema = apiModelIdProviderModelSchema.extend({
 	zaiApiLine: z.union([z.literal("china"), z.literal("international")]).optional(),
 })
 
+const fireworksSchema = apiModelIdProviderModelSchema.extend({
+	fireworksApiKey: z.string().optional(),
+})
+
 const defaultSchema = z.object({
 	apiProvider: z.undefined(),
 })
@@ -297,6 +302,7 @@ export const providerSettingsSchemaDiscriminated = z.discriminatedUnion("apiProv
 	cerebrasSchema.merge(z.object({ apiProvider: z.literal("cerebras") })),
 	sambaNovaSchema.merge(z.object({ apiProvider: z.literal("sambanova") })),
 	zaiSchema.merge(z.object({ apiProvider: z.literal("zai") })),
+	fireworksSchema.merge(z.object({ apiProvider: z.literal("fireworks") })),
 	defaultSchema,
 ])
 
@@ -331,6 +337,7 @@ export const providerSettingsSchema = z.object({
 	...cerebrasSchema.shape,
 	...sambaNovaSchema.shape,
 	...zaiSchema.shape,
+	...fireworksSchema.shape,
 	...codebaseIndexProviderSchema.shape,
 })
 

packages/types/src/providers/fireworks.ts

@@ -0,0 +1,61 @@
+import type { ModelInfo } from "../model.js"
+
+export type FireworksModelId =
+	| "accounts/fireworks/models/kimi-k2-instruct"
+	| "accounts/fireworks/models/qwen3-235b-a22b-instruct-2507"
+	| "accounts/fireworks/models/qwen3-coder-480b-a35b-instruct"
+	| "accounts/fireworks/models/deepseek-r1-0528"
+	| "accounts/fireworks/models/deepseek-v3"
+
+export const fireworksDefaultModelId: FireworksModelId = "accounts/fireworks/models/kimi-k2-instruct"
+
+export const fireworksModels = {
+	"accounts/fireworks/models/kimi-k2-instruct": {
+		maxTokens: 16384,
+		contextWindow: 128000,
+		supportsImages: false,
+		supportsPromptCache: false,
+		inputPrice: 0.6,
+		outputPrice: 2.5,
+		description:
+			"Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters. Trained with the Muon optimizer, Kimi K2 achieves exceptional performance across frontier knowledge, reasoning, and coding tasks while being meticulously optimized for agentic capabilities.",
+	},
+	"accounts/fireworks/models/qwen3-235b-a22b-instruct-2507": {
+		maxTokens: 32768,
+		contextWindow: 256000,
+		supportsImages: false,
+		supportsPromptCache: false,
+		inputPrice: 0.22,
+		outputPrice: 0.88,
+		description: "Latest Qwen3 thinking model, competitive against the best closed source models in Jul 2025.",
+	},
+	"accounts/fireworks/models/qwen3-coder-480b-a35b-instruct": {
+		maxTokens: 32768,
+		contextWindow: 256000,
+		supportsImages: false,
+		supportsPromptCache: false,
+		inputPrice: 0.45,
+		outputPrice: 1.8,
+		description: "Qwen3's most agentic code model to date.",
+	},
+	"accounts/fireworks/models/deepseek-r1-0528": {
+		maxTokens: 20480,
+		contextWindow: 160000,
+		supportsImages: false,
+		supportsPromptCache: false,
+		inputPrice: 3,
+		outputPrice: 8,
+		description:
+			"05/28 updated checkpoint of Deepseek R1. Its overall performance is now approaching that of leading models, such as O3 and Gemini 2.5 Pro. Compared to the previous version, the upgraded model shows significant improvements in handling complex reasoning tasks, and this version also offers a reduced hallucination rate, enhanced support for function calling, and better experience for vibe coding. Note that fine-tuning for this model is only available through contacting fireworks at https://fireworks.ai/company/contact-us.",
+	},
+	"accounts/fireworks/models/deepseek-v3": {
+		maxTokens: 16384,
+		contextWindow: 128000,
+		supportsImages: false,
+		supportsPromptCache: false,
+		inputPrice: 0.9,
+		outputPrice: 0.9,
+		description:
+			"A strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token from Deepseek. Note that fine-tuning for this model is only available through contacting fireworks at https://fireworks.ai/company/contact-us.",
+	},
+} as const satisfies Record<string, ModelInfo>

packages/types/src/providers/index.ts

@@ -23,3 +23,4 @@ export * from "./vscode-llm.js"
 export * from "./xai.js"
 export * from "./doubao.js"
 export * from "./zai.js"
+export * from "./fireworks.js"

src/api/index.ts

@@ -34,6 +34,7 @@ import {
 	SambaNovaHandler,
 	DoubaoHandler,
 	ZAiHandler,
+	FireworksHandler,
 } from "./providers"
 
 export interface SingleCompletionHandler {
@@ -127,6 +128,8 @@ export function buildApiHandler(configuration: ProviderSettings): ApiHandler {
 			return new SambaNovaHandler(options)
 		case "zai":
 			return new ZAiHandler(options)
+		case "fireworks":
+			return new FireworksHandler(options)
 		default:
 			apiProvider satisfies "gemini-cli" | undefined
 			return new AnthropicHandler(options)