feat: Add Kimi K2 0905 model to Groq, Moonshot, and Fireworks providers

- Added Kimi K2 0905 model with 256K context window
- Updated default models to use the new version
- Added prompt caching support where applicable
- Updated tests to reflect new model configurations

Author: daniel-lxs

packages/types/src/providers/fireworks.ts

@@ -2,6 +2,7 @@ import type { ModelInfo } from "../model.js"
 
 export type FireworksModelId =
 	| "accounts/fireworks/models/kimi-k2-instruct"
+	| "accounts/fireworks/models/kimi-k2-instruct-0905"
 	| "accounts/fireworks/models/qwen3-235b-a22b-instruct-2507"
 	| "accounts/fireworks/models/qwen3-coder-480b-a35b-instruct"
 	| "accounts/fireworks/models/deepseek-r1-0528"
@@ -12,9 +13,20 @@ export type FireworksModelId =
 	| "accounts/fireworks/models/gpt-oss-20b"
 	| "accounts/fireworks/models/gpt-oss-120b"
 
-export const fireworksDefaultModelId: FireworksModelId = "accounts/fireworks/models/kimi-k2-instruct"
+export const fireworksDefaultModelId: FireworksModelId = "accounts/fireworks/models/kimi-k2-instruct-0905"
 
 export const fireworksModels = {
+	"accounts/fireworks/models/kimi-k2-instruct-0905": {
+		maxTokens: 16384,
+		contextWindow: 262144,
+		supportsImages: false,
+		supportsPromptCache: true,
+		inputPrice: 0.6,
+		outputPrice: 2.5,
+		cacheReadsPrice: 0.15,
+		description:
+			"Kimi K2 model gets a new version update: Agentic coding: more accurate, better generalization across scaffolds. Frontend coding: improved aesthetics and functionalities on web, 3d, and other tasks. Context length: extended from 128k to 256k, providing better long-horizon support.",
+	},
 	"accounts/fireworks/models/kimi-k2-instruct": {
 		maxTokens: 16384,
 		contextWindow: 128000,

packages/types/src/providers/groq.ts

@@ -11,10 +11,11 @@ export type GroqModelId =
 	| "qwen/qwen3-32b"
 	| "deepseek-r1-distill-llama-70b"
 	| "moonshotai/kimi-k2-instruct"
+	| "moonshotai/kimi-k2-instruct-0905"
 	| "openai/gpt-oss-120b"
 	| "openai/gpt-oss-20b"
 
-export const groqDefaultModelId: GroqModelId = "llama-3.3-70b-versatile" // Defaulting to Llama3 70B Versatile
+export const groqDefaultModelId: GroqModelId = "moonshotai/kimi-k2-instruct-0905"
 
 export const groqModels = {
 	// Models based on API response: https://api.groq.com/openai/v1/models
@@ -100,6 +101,17 @@ export const groqModels = {
 		cacheReadsPrice: 0.5, // 50% discount for cached input tokens
 		description: "Moonshot AI Kimi K2 Instruct 1T model, 128K context.",
 	},
+	"moonshotai/kimi-k2-instruct-0905": {
+		maxTokens: 16384,
+		contextWindow: 262144,
+		supportsImages: false,
+		supportsPromptCache: true,
+		inputPrice: 0.6,
+		outputPrice: 2.5,
+		cacheReadsPrice: 0.15,
+		description:
+			"Kimi K2 model gets a new version update: Agentic coding: more accurate, better generalization across scaffolds. Frontend coding: improved aesthetics and functionalities on web, 3d, and other tasks. Context length: extended from 128k to 256k, providing better long-horizon support.",
+	},
 	"openai/gpt-oss-120b": {
 		maxTokens: 32766,
 		contextWindow: 131072,

packages/types/src/providers/moonshot.ts

@@ -18,15 +18,15 @@ export const moonshotModels = {
 		description: `Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters.`,
 	},
 	"kimi-k2-0905-preview": {
-		maxTokens: 32_000,
-		contextWindow: 262_144,
+		maxTokens: 16384,
+		contextWindow: 262144,
 		supportsImages: false,
 		supportsPromptCache: true,
-		inputPrice: 0.6, // $0.60 per million tokens (cache miss)
-		outputPrice: 2.5, // $2.50 per million tokens
-		cacheWritesPrice: 0, // $0 per million tokens (cache miss)
-		cacheReadsPrice: 0.15, // $0.15 per million tokens (cache hit)
-		description: `Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters.`,
+		inputPrice: 0.6,
+		outputPrice: 2.5,
+		cacheReadsPrice: 0.15,
+		description:
+			"Kimi K2 model gets a new version update: Agentic coding: more accurate, better generalization across scaffolds. Frontend coding: improved aesthetics and functionalities on web, 3d, and other tasks. Context length: extended from 128k to 256k, providing better long-horizon support.",
 	},
 	"kimi-k2-turbo-preview": {
 		maxTokens: 32_000,

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

@@ -148,8 +148,8 @@ describe("MoonshotHandler", () => {
 			const model = handler.getModel()
 			expect(model.id).toBe(mockOptions.apiModelId)
 			expect(model.info).toBeDefined()
-			expect(model.info.maxTokens).toBe(32_000)
-			expect(model.info.contextWindow).toBe(262_144)
+			expect(model.info.maxTokens).toBe(16384)
+			expect(model.info.contextWindow).toBe(262144)
 			expect(model.info.supportsImages).toBe(false)
 			expect(model.info.supportsPromptCache).toBe(true) // Should be true now
 		})