feat: add DeepSeek V4 model support with thinking/reasoning_content handling

packages/types/src/providers/deepseek.ts

@@ -32,6 +32,30 @@ 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-flash": {
+		maxTokens: 8192,
+		contextWindow: 128_000,
+		supportsImages: false,
+		supportsPromptCache: true,
+		preserveReasoning: true,
+		inputPrice: 0.28,
+		outputPrice: 0.42,
+		cacheWritesPrice: 0.28,
+		cacheReadsPrice: 0.028,
+		description: `DeepSeek V4 Flash with thinking/reasoning support. Requires reasoning_content to be passed back during tool call sequences.`,
+	},
+	"deepseek-v4-pro": {
+		maxTokens: 8192,
+		contextWindow: 128_000,
+		supportsImages: false,
+		supportsPromptCache: true,
+		preserveReasoning: true,
+		inputPrice: 0.28,
+		outputPrice: 0.42,
+		cacheWritesPrice: 0.28,
+		cacheReadsPrice: 0.028,
+		description: `DeepSeek V4 Pro with thinking/reasoning support. Requires reasoning_content to be passed back during tool call sequences.`,
+	},
 } 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
@@ -240,6 +241,24 @@ describe("DeepSeekHandler", () => {
 			expect((model.info as ModelInfo).preserveReasoning).toBe(true)
 		})
 
+		it("should have preserveReasoning enabled for deepseek-v4-flash", () => {
+			const handlerV4Flash = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-flash",
+			})
+			const model = handlerV4Flash.getModel()
+			expect((model.info as ModelInfo).preserveReasoning).toBe(true)
+		})
+
+		it("should have preserveReasoning enabled for deepseek-v4-pro", () => {
+			const handlerV4Pro = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-pro",
+			})
+			const model = handlerV4Pro.getModel()
+			expect((model.info as ModelInfo).preserveReasoning).toBe(true)
+		})
+
 		it("should NOT have preserveReasoning enabled for deepseek-chat", () => {
 			// deepseek-chat doesn't use thinking mode, so no need to preserve reasoning
 			const model = handler.getModel()
@@ -459,6 +478,44 @@ describe("DeepSeekHandler", () => {
 			)
 		})
 
+		it("should pass thinking parameter for deepseek-v4-flash model", async () => {
+			const v4FlashHandler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-flash",
+			})
+
+			const stream = v4FlashHandler.createMessage(systemPrompt, messages)
+			for await (const _chunk of stream) {
+				// Consume the stream
+			}
+
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					thinking: { type: "enabled" },
+				}),
+				{},
+			)
+		})
+
+		it("should pass thinking parameter for deepseek-v4-pro model", async () => {
+			const v4ProHandler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-pro",
+			})
+
+			const stream = v4ProHandler.createMessage(systemPrompt, messages)
+			for await (const _chunk of stream) {
+				// Consume the stream
+			}
+
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					thinking: { type: "enabled" },
+				}),
+				{},
+			)
+		})
+
 		it("should NOT pass thinking parameter for deepseek-chat model", async () => {
 			const chatHandler = new DeepSeekHandler({
 				...mockOptions,

src/api/providers/deepseek.ts

@@ -55,8 +55,9 @@ 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 by looking at the model info's preserveReasoning flag.
+		// This covers deepseek-reasoner and newer models like deepseek-v4-flash/deepseek-v4-pro.
+		const isThinkingModel = "preserveReasoning" in modelInfo && modelInfo.preserveReasoning === true
 
 		// 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

src/api/providers/openai.ts

@@ -89,7 +89,8 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 		const modelId = this.options.openAiModelId ?? ""
 		const enabledR1Format = this.options.openAiR1FormatEnabled ?? false
 		const isAzureAiInference = this._isAzureAiInference(modelUrl)
-		const deepseekReasoner = modelId.includes("deepseek-reasoner") || enabledR1Format
+		const deepseekReasoner =
+			modelId.includes("deepseek-reasoner") || modelId.includes("deepseek-v4") || enabledR1Format
 
 		if (modelId.includes("o1") || modelId.includes("o3") || modelId.includes("o4")) {
 			yield* this.handleO3FamilyMessage(modelId, systemPrompt, messages, metadata)