feat(models): add Baidu Qianfan International

.env.example

@@ -246,6 +246,9 @@ LLM_FIREWORKS_BASE_URL=https://api.fireworks.ai/inference
 # RanoAI
 LLM_RANOAI_API_KEY=your_ranoai_key_here
 
+# Baidu (Qianfan)
+LLM_BAIDU_API_KEY=your_baidu_qianfan_key_here
+
 # Google Vertex AI (Gemini models)
 LLM_GOOGLE_VERTEX_API_KEY=your_google_vertex_api_key_here
 LLM_GOOGLE_CLOUD_PROJECT=your_google_cloud_project_id

.env.unified.example

@@ -99,6 +99,9 @@ LLM_SCX_AI_API_KEY=your_scx_ai_key_here
 # RanoAI
 LLM_RANOAI_API_KEY=your_ranoai_key_here
 
+# Baidu (Qianfan)
+LLM_BAIDU_API_KEY=your_baidu_qianfan_key_here
+
 # Mistral
 LLM_MISTRAL_API_KEY=your_mistral_key_here
 

.github/workflows/e2e.yml

@@ -145,6 +145,7 @@ jobs:
           LLM_GONKA_24_API_KEY: ${{ secrets.LLM_GONKA_24_API_KEY }}
           LLM_FIREWORKS_API_KEY: ${{ secrets.LLM_FIREWORKS_API_KEY }}
           LLM_RANOAI_API_KEY: ${{ secrets.LLM_RANOAI_API_KEY }}
+          LLM_BAIDU_API_KEY: ${{ secrets.LLM_BAIDU_API_KEY }}
 
       - name: Upload shard results
         if: always()

apps/gateway/src/chat/tools/transform-streaming-to-openai.ts

@@ -1439,6 +1439,7 @@ export function transformStreamingToOpenai(
 		case "runware":
 		case "gonka24":
 		case "ranoai":
+		case "baidu":
 		case "granite":
 		case "tundra":
 		case "xiaomi":

apps/gateway/src/lib/costs.spec.ts

@@ -736,6 +736,28 @@ describe("calculateCosts", () => {
 		expect(result.completionTokens).toBe(330);
 	});
 
+	it("should not double-bill Baidu reasoning tokens", async () => {
+		// Qianfan reports reasoning in completion_tokens_details while already
+		// counting it inside completion_tokens: a thinking-only reply comes back
+		// as completion_tokens === reasoning_tokens, so adding it again would
+		// double the billed output.
+		const result = await calculateCosts(
+			"ernie-5.0",
+			"baidu",
+			null,
+			1000,
+			400, // completionTokens already includes the 400 reasoning tokens
+			null,
+			undefined,
+			400,
+		);
+
+		// inputPrice 1.4e-6, outputPrice 5.6e-6.
+		expect(result.inputCost).toBeCloseTo(0.0014, 10);
+		expect(result.outputCost).toBeCloseTo(0.00224, 10); // 400 * 5.6e-6, not 800
+		expect(result.completionTokens).toBe(400);
+	});
+
 	it("should handle null reasoning tokens gracefully", async () => {
 		const result = await calculateCosts(
 			"gemini-2.5-pro",

apps/gateway/src/lib/costs.ts

@@ -738,8 +738,9 @@ export async function calculateCosts(
 	// reports reasoning in `completion_tokens_details` (which the streaming
 	// transform hoists to a top-level `reasoning_tokens`) while already counting
 	// it inside `completion_tokens`, so adding it again would roughly double the
-	// billed output on reasoning requests. For remaining providers, add
-	// reasoning separately.
+	// billed output on reasoning requests. Baidu's Qianfan reports the same way
+	// (a thinking-only reply returns completion_tokens === reasoning_tokens).
+	// For remaining providers, add reasoning separately.
 	const completionIncludesReasoning =
 		provider === "google-ai-studio" ||
 		provider === "glacier" ||
@@ -751,6 +752,7 @@ export async function calculateCosts(
 		provider === "sakana" ||
 		provider === "meta" ||
 		provider === "ranoai" ||
+		provider === "baidu" ||
 		provider === "aws-mantle";
 	const totalOutputTokens = completionIncludesReasoning
 		? calculatedCompletionTokens

apps/gateway/src/models/models.spec.ts

@@ -170,14 +170,15 @@ describe("Models API", () => {
 			);
 		}
 
-		// deepseek-v3.2 has a deactivated Nebius mapping declaring 32768; the
-		// advertised bound must come from the mappings that can still serve
-		// (65536), not from limits that no longer apply.
-		const deepseek = json.data.find(
-			(m: { id: string }) => m.id === "deepseek-v3.2",
-		);
-		expect(deepseek).toBeDefined();
-		expect(deepseek.max_output).toBe(65536);
+		// kimi-k2 has a deactivated mapping declaring 8192; the advertised bound
+		// must come from the mappings that can still serve (131072), not from
+		// limits that no longer apply. deepseek-v3.2 used to pin this case, but
+		// Qianfan serves it with a genuine 32768 output cap, which matches the
+		// deactivated Nebius limit and left the assertion unable to tell the two
+		// apart.
+		const kimi = json.data.find((m: { id: string }) => m.id === "kimi-k2");
+		expect(kimi).toBeDefined();
+		expect(kimi.max_output).toBe(131072);
 	});
 
 	test("GET /v1/models exposes reasoning_efforts on provider mappings that define them", async () => {

apps/ui/src/components/provider-keys/provider-logo.ts

@@ -48,6 +48,7 @@ export const providerLogoUrls: Partial<
 	fireworks: ProviderIcons.fireworks,
 	runware: ProviderIcons.runware,
 	ranoai: ProviderIcons.ranoai,
+	baidu: ProviderIcons.baidu,
 };
 
 export const getProviderLogoDarkModeClasses = () => {

infra/helm/llmgateway/values.yaml

@@ -466,6 +466,7 @@ llmProviders:
   # LLM_TOGETHER_AI_API_KEY: ""
   # LLM_SCX_AI_API_KEY: ""
   # LLM_RANOAI_API_KEY: ""
+  # LLM_BAIDU_API_KEY: ""
   # LLM_ALIBABA_API_KEY: ""
   # LLM_NEBIUS_API_KEY: ""
   # LLM_CEREBRAS_API_KEY: ""

packages/actions/src/get-provider-endpoint.ts

@@ -188,6 +188,7 @@ const PROVIDER_DEFAULT_BASE_URLS: Partial<Record<ProviderId, string>> = {
 	gonka24: "https://api.gonka24.com",
 	fireworks: "https://api.fireworks.ai/inference",
 	ranoai: "https://api.ranoai.com",
+	baidu: "https://qianfan.baidubce.com",
 };
 
 export function getProviderDefaultBaseUrl(
@@ -952,6 +953,9 @@ export function getProviderEndpoint(
 			}
 			return `${url}/v1/chat/completions`;
 		}
+		// Qianfan's OpenAI-compatible surface is versioned v2, not v1.
+		case "baidu":
+			return `${url}/v2/chat/completions`;
 		case "deepseek":
 		case "moonshot":
 		case "nebius":

packages/models/src/models.ts

@@ -2,6 +2,7 @@ import { alibabaModels } from "./models/alibaba.js";
 import { anthropicModels } from "./models/anthropic.js";
 import { atlascloudModels } from "./models/atlascloud.js";
 import { baaiModels } from "./models/baai.js";
+import { baiduModels } from "./models/baidu.js";
 import { bytedanceModels } from "./models/bytedance.js";
 import { deepseekModels } from "./models/deepseek.js";
 import { elevenlabsModels } from "./models/elevenlabs.js";
@@ -767,6 +768,7 @@ export const models = [
 	...alibabaModels,
 	...atlascloudModels,
 	...baaiModels,
+	...baiduModels,
 	...bytedanceModels,
 	...nousresearchModels,
 	...reveModels,

packages/models/src/models/baidu.ts

@@ -0,0 +1,31 @@
+import type { ModelDefinition } from "@/models.js";
+
+export const baiduModels = [
+	{
+		id: "ernie-5.0",
+		name: "ERNIE 5.0",
+		description:
+			"Baidu's flagship natively multimodal model with a 248k context window, accepting text, image, and video input and thinking before it answers.",
+		family: "baidu",
+		releasedAt: new Date("2026-03-05"),
+		providers: [
+			{
+				providerId: "baidu" as const,
+				externalId: "ernie-5.0",
+				inputPrice: "1.4e-6",
+				outputPrice: "5.6e-6",
+				requestPrice: "0",
+				contextSize: 248832,
+				maxOutput: 126976,
+				streaming: true,
+				reasoning: true,
+				vision: true,
+				tools: true,
+				// jsonOutput is left off across every Qianfan mapping until
+				// response_format is confirmed against a funded account: an
+				// unsupported claim 400s the request, while omitting it only
+				// keeps JSON-mode traffic on another provider.
+			},
+		],
+	},
+] as const satisfies ModelDefinition[];

packages/models/src/models/deepseek.ts

@@ -248,6 +248,19 @@ export const deepseekModels = [
 				tools: true,
 				jsonOutput: true,
 			},
+			{
+				providerId: "baidu",
+				externalId: "deepseek-v3.2",
+				inputPrice: "0.28e-6",
+				cachedInputPrice: "0.028e-6",
+				outputPrice: "0.42e-6",
+				requestPrice: "0",
+				contextSize: 131072,
+				maxOutput: 32768,
+				streaming: true,
+				vision: false,
+				tools: true,
+			},
 		],
 	},
 	{
@@ -487,6 +500,20 @@ export const deepseekModels = [
 				jsonOutput: true,
 				jsonOutputSchema: true,
 			},
+			{
+				providerId: "baidu",
+				externalId: "deepseek-v4-pro",
+				inputPrice: "1.69e-6",
+				cachedInputPrice: "0.14e-6",
+				outputPrice: "3.38e-6",
+				requestPrice: "0",
+				contextSize: 1441792,
+				maxOutput: 393216,
+				streaming: true,
+				reasoning: true,
+				vision: false,
+				tools: true,
+			},
 		],
 	},
 	{
@@ -774,6 +801,20 @@ export const deepseekModels = [
 				supportedToolChoices: ["auto", "none"],
 				jsonOutput: true,
 			},
+			{
+				providerId: "baidu",
+				externalId: "deepseek-v4-flash",
+				inputPrice: "0.14e-6",
+				cachedInputPrice: "0.028e-6",
+				outputPrice: "0.28e-6",
+				requestPrice: "0",
+				contextSize: 1441792,
+				maxOutput: 393216,
+				streaming: true,
+				reasoning: true,
+				vision: false,
+				tools: true,
+			},
 		],
 	},
 ] as const satisfies ModelDefinition[];

packages/models/src/models/moonshot.ts

@@ -563,6 +563,23 @@ export const moonshotModels = [
 				// normalize it defensively in both modes.
 				healStreamingJsonOutput: true,
 			},
+			{
+				providerId: "baidu",
+				externalId: "kimi-k2.6",
+				// Mirrors the Moonshot mapping: the model rejects forced
+				// tool_choice while thinking is on.
+				supportedToolChoices: ["auto", "none"],
+				inputPrice: "0.95e-6",
+				cachedInputPrice: "0.16e-6",
+				outputPrice: "4e-6",
+				requestPrice: "0",
+				contextSize: 524288,
+				maxOutput: 262144,
+				streaming: true,
+				reasoning: true,
+				vision: true,
+				tools: true,
+			},
 		],
 	},
 	{

packages/models/src/models/zai.ts

@@ -205,6 +205,20 @@ export const zaiModels = [
 				tools: true,
 				jsonOutput: true,
 			},
+			{
+				providerId: "baidu",
+				externalId: "glm-5.2",
+				inputPrice: "1.4e-6",
+				cachedInputPrice: "0.26e-6",
+				outputPrice: "4.4e-6",
+				requestPrice: "0",
+				contextSize: 1179648,
+				maxOutput: 131072,
+				streaming: true,
+				reasoning: true,
+				vision: false,
+				tools: true,
+			},
 		],
 	},
 	{
@@ -350,6 +364,20 @@ export const zaiModels = [
 				tools: true,
 				jsonOutput: true,
 			},
+			{
+				providerId: "baidu",
+				externalId: "glm-5.1",
+				inputPrice: "1.4e-6",
+				cachedInputPrice: "0.26e-6",
+				outputPrice: "4.4e-6",
+				requestPrice: "0",
+				contextSize: 333824,
+				maxOutput: 131072,
+				streaming: true,
+				reasoning: true,
+				vision: false,
+				tools: true,
+			},
 		],
 	},
 	{
@@ -506,6 +534,20 @@ export const zaiModels = [
 				tools: true,
 				jsonOutput: true,
 			},
+			{
+				providerId: "baidu",
+				externalId: "glm-5",
+				inputPrice: "1e-6",
+				cachedInputPrice: "0.2e-6",
+				outputPrice: "3.2e-6",
+				requestPrice: "0",
+				contextSize: 335872,
+				maxOutput: 131072,
+				streaming: true,
+				reasoning: true,
+				vision: false,
+				tools: true,
+			},
 		],
 	},
 	{

packages/models/src/providers.ts

@@ -1202,6 +1202,35 @@ export const providers: ProviderDefinition[] = [
 		},
 		priority: 1.2,
 	},
+	{
+		id: "baidu",
+		name: "Baidu",
+		description:
+			"Baidu's Qianfan platform serving ERNIE alongside DeepSeek, GLM, and Kimi models",
+		env: {
+			required: {
+				apiKey: "LLM_BAIDU_API_KEY",
+			},
+		},
+		streaming: true,
+		cancellation: true,
+		color: "#2932E1",
+		website: "https://cloud.baidu.com/product/qianfan",
+		statusPageUrl: null,
+		announcement: null,
+		termsUrl: "https://cloud.baidu.com/doc/Agreements/s/yjwvy1x03",
+		privacyPolicyUrl: "https://cloud.baidu.com/doc/Agreements/s/Plr0fi68q",
+		headquarters: "CN",
+		// Qianfan publishes no API training / prompt logging commitment we can
+		// point at, so every attribute stays unknown and fails closed under a
+		// compliance policy rather than claiming a guarantee Baidu never made.
+		dataPolicy: {
+			apiTraining: null,
+			consumerTraining: null,
+			promptLogging: null,
+			retentionPeriod: null,
+		},
+	},
 	{
 		id: "perplexity",
 		name: "Perplexity",