feat: add DeepSeek V4 model support and fix thinking mode detection

packages/types/src/providers/deepseek.ts

@@ -32,6 +32,19 @@ 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 models - https://api-docs.deepseek.com/quick_start/pricing
+	"deepseek-v4-0324": {
+		maxTokens: 16_384, // 16K max output
+		contextWindow: 128_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: 0.219, // $0.219 per million tokens (cache hit)
+		description: `DeepSeek-V4-0324 is the latest flagship reasoning model with significantly improved performance across math, code, and complex reasoning tasks. Features enhanced thinking mode with interleaved reasoning and supports vision, JSON output, tool calls, and extended context.`,
+	},
 } 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
@@ -247,6 +248,29 @@ describe("DeepSeekHandler", () => {
 			expect((model.info as ModelInfo).preserveReasoning).toBeUndefined()
 		})
 
+		it("should return correct model info for deepseek-v4-0324", () => {
+			const handlerWithV4 = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-0324",
+			})
+			const model = handlerWithV4.getModel()
+			expect(model.id).toBe("deepseek-v4-0324")
+			expect(model.info).toBeDefined()
+			expect(model.info.maxTokens).toBe(16_384)
+			expect(model.info.contextWindow).toBe(128_000)
+			expect(model.info.supportsImages).toBe(true)
+			expect(model.info.supportsPromptCache).toBe(true)
+		})
+
+		it("should have preserveReasoning enabled for deepseek-v4-0324", () => {
+			const handlerWithV4 = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-0324",
+			})
+			const model = handlerWithV4.getModel()
+			expect((model.info as ModelInfo).preserveReasoning).toBe(true)
+		})
+
 		it("should return provided model ID with default model info if model does not exist", () => {
 			const handlerWithInvalidModel = new DeepSeekHandler({
 				...mockOptions,
@@ -475,6 +499,44 @@ describe("DeepSeekHandler", () => {
 			expect(callArgs.thinking).toBeUndefined()
 		})
 
+		it("should pass thinking parameter for deepseek-v4-0324 model", async () => {
+			const v4Handler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-0324",
+			})
+
+			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 for v4 model
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					thinking: { type: "enabled" },
+				}),
+				{},
+			)
+		})
+
+		it("should handle reasoning_content in streaming responses for deepseek-v4-0324", async () => {
+			const v4Handler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-0324",
+			})
+
+			const stream = v4Handler.createMessage(systemPrompt, messages)
+			const chunks: any[] = []
+			for await (const chunk of stream) {
+				chunks.push(chunk)
+			}
+
+			// Should have reasoning chunks since v4 model has preserveReasoning
+			const reasoningChunks = chunks.filter((chunk) => chunk.type === "reasoning")
+			expect(reasoningChunks.length).toBeGreaterThan(0)
+			expect(reasoningChunks[0].text).toBe("Let me think about this...")
+		})
+
 		it("should handle tool calls with reasoning_content", async () => {
 			const reasonerHandler = new DeepSeekHandler({
 				...mockOptions,

src/api/providers/deepseek.ts

@@ -6,6 +6,7 @@ import {
 	deepSeekDefaultModelId,
 	DEEP_SEEK_DEFAULT_TEMPERATURE,
 	OPENAI_AZURE_AI_INFERENCE_PATH,
+	type ModelInfo,
 } from "@roo-code/types"
 
 import type { ApiHandlerOptions } from "../../shared/api"
@@ -55,8 +56,9 @@ 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 via the preserveReasoning flag
+		// This covers deepseek-reasoner and newer v4 models that support thinking mode
+		const isThinkingModel = (modelInfo as 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