feat: add deepseek-v4-pro and deepseek-v4-flash models

- Add deepseek-v4-pro and deepseek-v4-flash to DeepSeek model definitions
- Both models support thinking mode (preserveReasoning), vision, and prompt caching
- Update thinking mode detection to use modelInfo.preserveReasoning flag
  instead of hardcoded model name check, making it future-proof
- Add tests for new v4 model info and thinking mode behavior

Addresses #12174

Author: roomote

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-pro": {
+		maxTokens: 16_384, // 16K max output
+		contextWindow: 128_000,
+		supportsImages: true,
+		supportsPromptCache: true,
+		preserveReasoning: true,
+		inputPrice: 2.0, // $2.00 per million tokens (cache miss)
+		outputPrice: 8.0, // $8.00 per million tokens
+		cacheWritesPrice: 2.0, // $2.00 per million tokens (cache miss)
+		cacheReadsPrice: 0.5, // $0.50 per million tokens (cache hit)
+		description: `DeepSeek V4 Pro is a flagship reasoning model with thinking capabilities, vision support, and enhanced tool use. Excels at complex reasoning, coding, and multi-step problem solving tasks.`,
+	},
+	"deepseek-v4-flash": {
+		maxTokens: 16_384, // 16K max output
+		contextWindow: 128_000,
+		supportsImages: true,
+		supportsPromptCache: true,
+		preserveReasoning: true,
+		inputPrice: 1.0, // $1.00 per million tokens (cache miss)
+		outputPrice: 4.0, // $4.00 per million tokens
+		cacheWritesPrice: 1.0, // $1.00 per million tokens (cache miss)
+		cacheReadsPrice: 0.25, // $0.25 per million tokens (cache hit)
+		description: `DeepSeek V4 Flash is a fast, cost-efficient reasoning model with thinking capabilities and vision support. Optimized for speed while maintaining strong performance across coding, reasoning, and general tasks.`,
+	},
 } as const satisfies Record<string, ModelInfo>
 
 // https://api-docs.deepseek.com/quick_start/parameter_settings

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

@@ -29,8 +29,10 @@ vi.mock("openai", () => {
 							}
 						}
 
-						// Check if this is a reasoning_content test by looking at model
-						const isReasonerModel = options.model?.includes("deepseek-reasoner")
+						// Check if this is a thinking model - matches models with preserveReasoning: true
+						// (deepseek-reasoner, deepseek-v4-pro, deepseek-v4-flash)
+						const isReasonerModel =
+							options.model?.includes("deepseek-reasoner") || options.model?.includes("deepseek-v4-")
 						const isToolCallTest = options.tools?.length > 0
 
 						// Return async iterator for streaming
@@ -122,7 +124,7 @@ vi.mock("openai", () => {
 import OpenAI from "openai"
 import type { Anthropic } from "@anthropic-ai/sdk"
 
-import { deepSeekDefaultModelId, DEEP_SEEK_DEFAULT_TEMPERATURE, type ModelInfo } from "@roo-code/types"
+import { deepSeekDefaultModelId, deepSeekModels, DEEP_SEEK_DEFAULT_TEMPERATURE, type ModelInfo } from "@roo-code/types"
 
 import type { ApiHandlerOptions } from "../../../shared/api"
 
@@ -247,6 +249,36 @@ describe("DeepSeekHandler", () => {
 			expect((model.info as ModelInfo).preserveReasoning).toBeUndefined()
 		})
 
+		it("should return correct model info for deepseek-v4-pro", () => {
+			const handlerWithV4Pro = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-pro",
+			})
+			const model = handlerWithV4Pro.getModel()
+			expect(model.id).toBe("deepseek-v4-pro")
+			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)
+			expect((model.info as ModelInfo).preserveReasoning).toBe(true)
+		})
+
+		it("should return correct model info for deepseek-v4-flash", () => {
+			const handlerWithV4Flash = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-flash",
+			})
+			const model = handlerWithV4Flash.getModel()
+			expect(model.id).toBe("deepseek-v4-flash")
+			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)
+			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 +507,61 @@ describe("DeepSeekHandler", () => {
 			expect(callArgs.thinking).toBeUndefined()
 		})
 
+		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 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 handle reasoning_content in streaming responses for deepseek-v4-pro", async () => {
+			const v4ProHandler = new DeepSeekHandler({
+				...mockOptions,
+				apiModelId: "deepseek-v4-pro",
+			})
+
+			const stream = v4ProHandler.createMessage(systemPrompt, messages)
+			const chunks: any[] = []
+			for await (const chunk of stream) {
+				chunks.push(chunk)
+			}
+
+			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 using the preserveReasoning flag from ModelInfo.
+		// This covers deepseek-reasoner, deepseek-v4-pro, deepseek-v4-flash, and any future thinking models.
+		const isThinkingModel = !!(modelInfo as ModelInfo).preserveReasoning
 
 		// 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