feat: add DeepSeek V4 Pro model support

packages/types/src/providers/deepseek.ts

@@ -32,6 +32,18 @@ export const deepSeekModels = {
 		cacheReadsPrice: 0.028, // $0.028 per million tokens (cache hit) - Updated Dec 9, 2025
 		description: `DeepSeek-V3.2 (Thinking Mode) achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. Supports Chain of Thought reasoning with up to 8K output tokens. Supports JSON output, tool calls, and chat prefix completion (beta).`,
 	},
+	"deepseek-v4-pro": {
+		maxTokens: 16_384, // 16K max output
+		contextWindow: 164_000,
+		supportsImages: true,
+		supportsPromptCache: true,
+		preserveReasoning: true,
+		inputPrice: 2.19, // $2.19 per million tokens (cache miss)
+		outputPrice: 8.87, // $8.87 per million tokens
+		cacheWritesPrice: 2.19, // $2.19 per million tokens (cache miss)
+		cacheReadsPrice: 0.55, // $0.55 per million tokens (cache hit)
+		description: `DeepSeek V4 Pro is a frontier reasoning model with advanced capabilities across math, code, and complex reasoning tasks. Features 164K context window, 16K max output, vision support, and enhanced tool calling with interleaved thinking mode.`,
+	},
 } as const satisfies Record<string, ModelInfo>
 
 // https://api-docs.deepseek.com/quick_start/parameter_settings

src/api/providers/__tests__/deepseek.spec.ts

@@ -30,7 +30,8 @@ vi.mock("openai", () => {
 						}
 
 						// Check if this is a reasoning_content test by looking at model
-						const isReasonerModel = options.model?.includes("deepseek-reasoner")
+						const isReasonerModel =
+							options.model?.includes("deepseek-reasoner") || options.model?.includes("deepseek-v4")
 						const isToolCallTest = options.tools?.length > 0
 
 						// Return async iterator for streaming
@@ -507,5 +508,69 @@ describe("DeepSeekHandler", () => {
 			expect(toolCallChunks.length).toBeGreaterThan(0)
 			expect(toolCallChunks[0].name).toBe("get_weather")
 		})
+
+		it("should handle reasoning_content in streaming responses for deepseek-v4-pro", async () => {
+			const v4Handler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-pro",
+			})
+
+			const stream = v4Handler.createMessage(systemPrompt, messages)
+			const chunks: any[] = []
+			for await (const chunk of stream) {
+				chunks.push(chunk)
+			}
+
+			// Should have reasoning chunks
+			const reasoningChunks = chunks.filter((chunk) => chunk.type === "reasoning")
+			expect(reasoningChunks.length).toBeGreaterThan(0)
+			expect(reasoningChunks[0].text).toBe("Let me think about this...")
+			expect(reasoningChunks[1].text).toBe(" I'll analyze step by step.")
+		})
+
+		it("should pass thinking parameter for deepseek-v4-pro model", async () => {
+			const v4Handler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-pro",
+			})
+
+			const stream = v4Handler.createMessage(systemPrompt, messages)
+			for await (const _chunk of stream) {
+				// Consume the stream
+			}
+
+			// Verify that the thinking parameter was passed to the API
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					thinking: { type: "enabled" },
+				}),
+				{}, // Empty path options for non-Azure URLs
+			)
+		})
+
+		it("should have preserveReasoning enabled for deepseek-v4-pro", () => {
+			const v4Handler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-pro",
+			})
+			const model = v4Handler.getModel()
+			expect((model.info as ModelInfo).preserveReasoning).toBe(true)
+		})
+	})
+
+	describe("deepseek-v4-pro model info", () => {
+		it("should return correct model info for deepseek-v4-pro", () => {
+			const v4Handler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-pro",
+			})
+			const model = v4Handler.getModel()
+			expect(model.id).toBe("deepseek-v4-pro")
+			expect(model.info).toBeDefined()
+			expect(model.info.maxTokens).toBe(16_384)
+			expect(model.info.contextWindow).toBe(164_000)
+			expect(model.info.supportsImages).toBe(true)
+			expect(model.info.supportsPromptCache).toBe(true)
+		})
 	})
 })

src/api/providers/deepseek.ts

@@ -55,8 +55,8 @@ export class DeepSeekHandler extends OpenAiHandler {
 		const modelId = this.options.apiModelId ?? deepSeekDefaultModelId
 		const { info: modelInfo } = this.getModel()
 
-		// Check if this is a thinking-enabled model (deepseek-reasoner)
-		const isThinkingModel = modelId.includes("deepseek-reasoner")
+		// Check if this is a thinking-enabled model (deepseek-reasoner or deepseek-v4-pro)
+		const isThinkingModel = modelId.includes("deepseek-reasoner") || modelId.includes("deepseek-v4")
 
 		// Convert messages to R1 format (merges consecutive same-role messages)
 		// This is required for DeepSeek which does not support successive messages with the same role