@tanstack/ai-groq 0.4.16 → 0.5.1

package/dist/esm/model-meta.js

@@ -48,7 +48,23 @@ const GROQ_CHAT_MODELS = [
   KIMI_K2_INSTRUCT_0905.name,
   QWEN3_32B.name
 ];
+const GROQ_TRANSCRIPTION_MODELS = [
+  "whisper-large-v3-turbo",
+  "whisper-large-v3"
+];
+const ORPHEUS_V1_ENGLISH = {
+  name: "canopylabs/orpheus-v1-english"
+};
+const ORPHEUS_ARABIC_SAUDI = {
+  name: "canopylabs/orpheus-arabic-saudi"
+};
+const GROQ_TTS_MODELS = [
+  ORPHEUS_V1_ENGLISH.name,
+  ORPHEUS_ARABIC_SAUDI.name
+];
 export {
-  GROQ_CHAT_MODELS
+  GROQ_CHAT_MODELS,
+  GROQ_TRANSCRIPTION_MODELS,
+  GROQ_TTS_MODELS
 };
 //# sourceMappingURL=model-meta.js.map

package/dist/esm/model-meta.js.map

@@ -1 +1 @@
-{"version":3,"file":"model-meta.js","sources":["../../src/model-meta.ts"],"sourcesContent":["import type { GroqTextProviderOptions } from './text/text-provider-options'\n\n/**\n * Internal metadata structure describing a Groq model's capabilities and pricing.\n */\ninterface ModelMeta<TProviderOptions = unknown> {\n  name: string\n  context_window?: number\n  max_completion_tokens?: number\n  pricing: {\n    input?: { normal: number; cached?: number }\n    output?: { normal: number }\n  }\n  supports: {\n    input: Array<'text' | 'image' | 'audio'>\n    output: Array<'text' | 'audio'>\n    endpoints: Array<'chat' | 'tts' | 'transcription' | 'batch'>\n\n    features: Array<\n      | 'streaming'\n      | 'tools'\n      | 'json_object'\n      | 'browser_search'\n      | 'code_execution'\n      | 'reasoning'\n      | 'content_moderation'\n      | 'json_schema'\n      | 'vision'\n    >\n    tools?: ReadonlyArray<never>\n  }\n  /**\n   * Type-level description of which provider options this model supports.\n   */\n  providerOptions?: TProviderOptions\n}\n\nconst LLAMA_3_3_70B_VERSATILE = {\n  name: 'llama-3.3-70b-versatile',\n  context_window: 131_072,\n  max_completion_tokens: 32_768,\n  pricing: {\n    input: {\n      normal: 0.59,\n    },\n    output: {\n      normal: 0.79,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'tools', 'json_object'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_4_MAVERICK_17B_128E_INSTRUCT = {\n  name: 'meta-llama/llama-4-maverick-17b-128e-instruct',\n  context_window: 131_072,\n  max_completion_tokens: 8_192,\n  pricing: {\n    input: {\n      normal: 0.2,\n    },\n    output: {\n      normal: 0.6,\n    },\n  },\n  supports: {\n    input: ['text', 'image'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'tools', 'json_object', 'json_schema', 'vision'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_4_SCOUT_17B_16E_INSTRUCT = {\n  name: 'meta-llama/llama-4-scout-17b-16e-instruct',\n  context_window: 131_072,\n  max_completion_tokens: 8_192,\n  pricing: {\n    input: {\n      normal: 0.05,\n    },\n    output: {\n      normal: 0.08,\n    },\n  },\n  supports: {\n    input: ['text', 'image'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'tools', 'json_object'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_GUARD_4_12B = {\n  name: 'meta-llama/llama-guard-4-12b',\n  context_window: 131_072,\n  max_completion_tokens: 1024,\n  pricing: {\n    input: {\n      normal: 0.2,\n    },\n    output: {\n      normal: 0.2,\n    },\n  },\n  supports: {\n    input: ['text', 'image'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'json_object', 'content_moderation', 'vision'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_PROMPT_GUARD_2_86M = {\n  name: 'meta-llama/llama-prompt-guard-2-86m',\n  context_window: 512,\n  max_completion_tokens: 512,\n  pricing: {\n    input: {\n      normal: 0.04,\n    },\n    output: {\n      normal: 0.04,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'content_moderation', 'json_object'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_3_1_8B_INSTANT = {\n  name: 'llama-3.1-8b-instant',\n  context_window: 131_072,\n  max_completion_tokens: 131_072,\n  pricing: {\n    input: {\n      normal: 0.05,\n    },\n    output: {\n      normal: 0.08,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'json_object', 'tools'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_PROMPT_GUARD_2_22M = {\n  name: 'meta-llama/llama-prompt-guard-2-22m',\n  context_window: 512,\n  max_completion_tokens: 512,\n  pricing: {\n    input: {\n      normal: 0.03,\n    },\n    output: {\n      normal: 0.03,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'content_moderation'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst GPT_OSS_120B = {\n  name: 'openai/gpt-oss-120b',\n  context_window: 131_072,\n  max_completion_tokens: 65_536,\n  pricing: {\n    input: {\n      normal: 0.15,\n      cached: 0.075,\n    },\n    output: {\n      normal: 0.6,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: [\n      'streaming',\n      'json_object',\n      'json_schema',\n      'tools',\n      'browser_search',\n      'code_execution',\n      'reasoning',\n    ],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst GPT_OSS_SAFEGUARD_20B = {\n  name: 'openai/gpt-oss-safeguard-20b',\n  context_window: 131_072,\n  max_completion_tokens: 65_536,\n  pricing: {\n    input: {\n      normal: 0.075,\n      cached: 0.037,\n    },\n    output: {\n      normal: 0.3,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: [\n      'streaming',\n      'tools',\n      'browser_search',\n      'code_execution',\n      'json_object',\n      'json_schema',\n      'reasoning',\n      'content_moderation',\n    ],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst GPT_OSS_20B = {\n  name: 'openai/gpt-oss-20b',\n  context_window: 131_072,\n  max_completion_tokens: 65_536,\n  pricing: {\n    input: {\n      normal: 0.075,\n      cached: 0.037,\n    },\n    output: {\n      normal: 0.3,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: [\n      'streaming',\n      'browser_search',\n      'code_execution',\n      'json_object',\n      'json_schema',\n      'reasoning',\n      'tools',\n    ],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst KIMI_K2_INSTRUCT_0905 = {\n  name: 'moonshotai/kimi-k2-instruct-0905',\n  context_window: 262_144,\n  max_completion_tokens: 16_384,\n  pricing: {\n    input: {\n      normal: 1,\n      cached: 0.5,\n    },\n    output: {\n      normal: 3,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'tools', 'json_object', 'json_schema'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst QWEN3_32B = {\n  name: 'qwen/qwen3-32b',\n  context_window: 131_072,\n  max_completion_tokens: 40_960,\n  pricing: {\n    input: {\n      normal: 0.29,\n    },\n    output: {\n      normal: 0.59,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'json_object', 'tools', 'reasoning'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\n/**\n * All supported Groq chat model identifiers.\n */\nexport const GROQ_CHAT_MODELS = [\n  LLAMA_3_1_8B_INSTANT.name,\n  LLAMA_3_3_70B_VERSATILE.name,\n  LLAMA_4_MAVERICK_17B_128E_INSTRUCT.name,\n  LLAMA_4_SCOUT_17B_16E_INSTRUCT.name,\n  LLAMA_GUARD_4_12B.name,\n  LLAMA_PROMPT_GUARD_2_86M.name,\n  LLAMA_PROMPT_GUARD_2_22M.name,\n  GPT_OSS_20B.name,\n  GPT_OSS_120B.name,\n  GPT_OSS_SAFEGUARD_20B.name,\n  KIMI_K2_INSTRUCT_0905.name,\n  QWEN3_32B.name,\n] as const\n\n/**\n * Union type of all supported Groq chat model names.\n */\nexport type GroqChatModels = (typeof GROQ_CHAT_MODELS)[number]\n\n/**\n * Type-only map from Groq chat model name to its supported input modalities.\n */\nexport type GroqModelInputModalitiesByName = {\n  [LLAMA_3_1_8B_INSTANT.name]: typeof LLAMA_3_1_8B_INSTANT.supports.input\n  [LLAMA_3_3_70B_VERSATILE.name]: typeof LLAMA_3_3_70B_VERSATILE.supports.input\n  [LLAMA_4_MAVERICK_17B_128E_INSTRUCT.name]: typeof LLAMA_4_MAVERICK_17B_128E_INSTRUCT.supports.input\n  [LLAMA_4_SCOUT_17B_16E_INSTRUCT.name]: typeof LLAMA_4_SCOUT_17B_16E_INSTRUCT.supports.input\n  [LLAMA_GUARD_4_12B.name]: typeof LLAMA_GUARD_4_12B.supports.input\n  [LLAMA_PROMPT_GUARD_2_86M.name]: typeof LLAMA_PROMPT_GUARD_2_86M.supports.input\n  [LLAMA_PROMPT_GUARD_2_22M.name]: typeof LLAMA_PROMPT_GUARD_2_22M.supports.input\n  [GPT_OSS_20B.name]: typeof GPT_OSS_20B.supports.input\n  [GPT_OSS_120B.name]: typeof GPT_OSS_120B.supports.input\n  [GPT_OSS_SAFEGUARD_20B.name]: typeof GPT_OSS_SAFEGUARD_20B.supports.input\n  [KIMI_K2_INSTRUCT_0905.name]: typeof KIMI_K2_INSTRUCT_0905.supports.input\n  [QWEN3_32B.name]: typeof QWEN3_32B.supports.input\n}\n\n/**\n * Type-only map from Groq chat model name to its provider options type.\n */\nexport type GroqChatModelProviderOptionsByName = {\n  [K in (typeof GROQ_CHAT_MODELS)[number]]: GroqTextProviderOptions\n}\n\n/**\n * Type-only map from Groq chat model name to its supported provider tools.\n * Groq exposes no provider-specific tool factories, so every model gets an\n * empty tuple. This ensures that passing an Anthropic/OpenAI ProviderTool to\n * a Groq adapter produces a compile-time type error.\n */\nexport type GroqChatModelToolCapabilitiesByName = {\n  [LLAMA_3_1_8B_INSTANT.name]: typeof LLAMA_3_1_8B_INSTANT.supports.tools\n  [LLAMA_3_3_70B_VERSATILE.name]: typeof LLAMA_3_3_70B_VERSATILE.supports.tools\n  [LLAMA_4_MAVERICK_17B_128E_INSTRUCT.name]: typeof LLAMA_4_MAVERICK_17B_128E_INSTRUCT.supports.tools\n  [LLAMA_4_SCOUT_17B_16E_INSTRUCT.name]: typeof LLAMA_4_SCOUT_17B_16E_INSTRUCT.supports.tools\n  [LLAMA_GUARD_4_12B.name]: typeof LLAMA_GUARD_4_12B.supports.tools\n  [LLAMA_PROMPT_GUARD_2_86M.name]: typeof LLAMA_PROMPT_GUARD_2_86M.supports.tools\n  [LLAMA_PROMPT_GUARD_2_22M.name]: typeof LLAMA_PROMPT_GUARD_2_22M.supports.tools\n  [GPT_OSS_20B.name]: typeof GPT_OSS_20B.supports.tools\n  [GPT_OSS_120B.name]: typeof GPT_OSS_120B.supports.tools\n  [GPT_OSS_SAFEGUARD_20B.name]: typeof GPT_OSS_SAFEGUARD_20B.supports.tools\n  [KIMI_K2_INSTRUCT_0905.name]: typeof KIMI_K2_INSTRUCT_0905.supports.tools\n  [QWEN3_32B.name]: typeof QWEN3_32B.supports.tools\n}\n\n/**\n * Resolves the provider options type for a specific Groq model.\n * Falls back to generic GroqTextProviderOptions for unknown models.\n */\nexport type ResolveProviderOptions<TModel extends string> =\n  TModel extends keyof GroqChatModelProviderOptionsByName\n    ? GroqChatModelProviderOptionsByName[TModel]\n    : GroqTextProviderOptions\n\n/**\n * Resolve input modalities for a specific model.\n * If the model has explicit modalities in the map, use those; otherwise use text only.\n */\nexport type ResolveInputModalities<TModel extends string> =\n  TModel extends keyof GroqModelInputModalitiesByName\n    ? GroqModelInputModalitiesByName[TModel]\n    : readonly ['text']\n"],"names":[],"mappings":"AAqCA,MAAM,0BAA0B;AAAA,EAC9B,MAAM;AAkBR;AAEA,MAAM,qCAAqC;AAAA,EACzC,MAAM;AAkBR;AAEA,MAAM,iCAAiC;AAAA,EACrC,MAAM;AAkBR;AAEA,MAAM,oBAAoB;AAAA,EACxB,MAAM;AAkBR;AAEA,MAAM,2BAA2B;AAAA,EAC/B,MAAM;AAkBR;AAEA,MAAM,uBAAuB;AAAA,EAC3B,MAAM;AAkBR;AAEA,MAAM,2BAA2B;AAAA,EAC/B,MAAM;AAkBR;AAEA,MAAM,eAAe;AAAA,EACnB,MAAM;AA2BR;AAEA,MAAM,wBAAwB;AAAA,EAC5B,MAAM;AA4BR;AAEA,MAAM,cAAc;AAAA,EAClB,MAAM;AA2BR;AAEA,MAAM,wBAAwB;AAAA,EAC5B,MAAM;AAmBR;AAEA,MAAM,YAAY;AAAA,EAChB,MAAM;AAkBR;AAKO,MAAM,mBAAmB;AAAA,EAC9B,qBAAqB;AAAA,EACrB,wBAAwB;AAAA,EACxB,mCAAmC;AAAA,EACnC,+BAA+B;AAAA,EAC/B,kBAAkB;AAAA,EAClB,yBAAyB;AAAA,EACzB,yBAAyB;AAAA,EACzB,YAAY;AAAA,EACZ,aAAa;AAAA,EACb,sBAAsB;AAAA,EACtB,sBAAsB;AAAA,EACtB,UAAU;AACZ;"}
+{"version":3,"file":"model-meta.js","sources":["../../src/model-meta.ts"],"sourcesContent":["import type { GroqTextProviderOptions } from './text/text-provider-options'\nimport type { GroqTTSProviderOptions } from './audio/tts-provider-options'\n\n/**\n * Internal metadata structure describing a Groq model's capabilities and pricing.\n */\ninterface ModelMeta<TProviderOptions = unknown> {\n  name: string\n  context_window?: number\n  max_completion_tokens?: number\n  pricing: {\n    input?: { normal: number; cached?: number }\n    output?: { normal: number }\n  }\n  supports: {\n    input: Array<'text' | 'image' | 'audio'>\n    output: Array<'text' | 'audio'>\n    endpoints: Array<'chat' | 'tts' | 'transcription' | 'batch'>\n\n    features: Array<\n      | 'streaming'\n      | 'tools'\n      | 'json_object'\n      | 'browser_search'\n      | 'code_execution'\n      | 'reasoning'\n      | 'content_moderation'\n      | 'json_schema'\n      | 'vision'\n    >\n    tools?: ReadonlyArray<never>\n  }\n  /**\n   * Type-level description of which provider options this model supports.\n   */\n  providerOptions?: TProviderOptions\n}\n\nconst LLAMA_3_3_70B_VERSATILE = {\n  name: 'llama-3.3-70b-versatile',\n  context_window: 131_072,\n  max_completion_tokens: 32_768,\n  pricing: {\n    input: {\n      normal: 0.59,\n    },\n    output: {\n      normal: 0.79,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'tools', 'json_object'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_4_MAVERICK_17B_128E_INSTRUCT = {\n  name: 'meta-llama/llama-4-maverick-17b-128e-instruct',\n  context_window: 131_072,\n  max_completion_tokens: 8_192,\n  pricing: {\n    input: {\n      normal: 0.2,\n    },\n    output: {\n      normal: 0.6,\n    },\n  },\n  supports: {\n    input: ['text', 'image'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'tools', 'json_object', 'json_schema', 'vision'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_4_SCOUT_17B_16E_INSTRUCT = {\n  name: 'meta-llama/llama-4-scout-17b-16e-instruct',\n  context_window: 131_072,\n  max_completion_tokens: 8_192,\n  pricing: {\n    input: {\n      normal: 0.05,\n    },\n    output: {\n      normal: 0.08,\n    },\n  },\n  supports: {\n    input: ['text', 'image'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'tools', 'json_object'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_GUARD_4_12B = {\n  name: 'meta-llama/llama-guard-4-12b',\n  context_window: 131_072,\n  max_completion_tokens: 1024,\n  pricing: {\n    input: {\n      normal: 0.2,\n    },\n    output: {\n      normal: 0.2,\n    },\n  },\n  supports: {\n    input: ['text', 'image'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'json_object', 'content_moderation', 'vision'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_PROMPT_GUARD_2_86M = {\n  name: 'meta-llama/llama-prompt-guard-2-86m',\n  context_window: 512,\n  max_completion_tokens: 512,\n  pricing: {\n    input: {\n      normal: 0.04,\n    },\n    output: {\n      normal: 0.04,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'content_moderation', 'json_object'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_3_1_8B_INSTANT = {\n  name: 'llama-3.1-8b-instant',\n  context_window: 131_072,\n  max_completion_tokens: 131_072,\n  pricing: {\n    input: {\n      normal: 0.05,\n    },\n    output: {\n      normal: 0.08,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'json_object', 'tools'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst LLAMA_PROMPT_GUARD_2_22M = {\n  name: 'meta-llama/llama-prompt-guard-2-22m',\n  context_window: 512,\n  max_completion_tokens: 512,\n  pricing: {\n    input: {\n      normal: 0.03,\n    },\n    output: {\n      normal: 0.03,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'content_moderation'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst GPT_OSS_120B = {\n  name: 'openai/gpt-oss-120b',\n  context_window: 131_072,\n  max_completion_tokens: 65_536,\n  pricing: {\n    input: {\n      normal: 0.15,\n      cached: 0.075,\n    },\n    output: {\n      normal: 0.6,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: [\n      'streaming',\n      'json_object',\n      'json_schema',\n      'tools',\n      'browser_search',\n      'code_execution',\n      'reasoning',\n    ],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst GPT_OSS_SAFEGUARD_20B = {\n  name: 'openai/gpt-oss-safeguard-20b',\n  context_window: 131_072,\n  max_completion_tokens: 65_536,\n  pricing: {\n    input: {\n      normal: 0.075,\n      cached: 0.037,\n    },\n    output: {\n      normal: 0.3,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: [\n      'streaming',\n      'tools',\n      'browser_search',\n      'code_execution',\n      'json_object',\n      'json_schema',\n      'reasoning',\n      'content_moderation',\n    ],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst GPT_OSS_20B = {\n  name: 'openai/gpt-oss-20b',\n  context_window: 131_072,\n  max_completion_tokens: 65_536,\n  pricing: {\n    input: {\n      normal: 0.075,\n      cached: 0.037,\n    },\n    output: {\n      normal: 0.3,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: [\n      'streaming',\n      'browser_search',\n      'code_execution',\n      'json_object',\n      'json_schema',\n      'reasoning',\n      'tools',\n    ],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst KIMI_K2_INSTRUCT_0905 = {\n  name: 'moonshotai/kimi-k2-instruct-0905',\n  context_window: 262_144,\n  max_completion_tokens: 16_384,\n  pricing: {\n    input: {\n      normal: 1,\n      cached: 0.5,\n    },\n    output: {\n      normal: 3,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'tools', 'json_object', 'json_schema'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\nconst QWEN3_32B = {\n  name: 'qwen/qwen3-32b',\n  context_window: 131_072,\n  max_completion_tokens: 40_960,\n  pricing: {\n    input: {\n      normal: 0.29,\n    },\n    output: {\n      normal: 0.59,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['text'],\n    endpoints: ['chat'],\n    features: ['streaming', 'json_object', 'tools', 'reasoning'],\n    tools: [] as const,\n  },\n} as const satisfies ModelMeta<GroqTextProviderOptions>\n\n/**\n * All supported Groq chat model identifiers.\n */\nexport const GROQ_CHAT_MODELS = [\n  LLAMA_3_1_8B_INSTANT.name,\n  LLAMA_3_3_70B_VERSATILE.name,\n  LLAMA_4_MAVERICK_17B_128E_INSTRUCT.name,\n  LLAMA_4_SCOUT_17B_16E_INSTRUCT.name,\n  LLAMA_GUARD_4_12B.name,\n  LLAMA_PROMPT_GUARD_2_86M.name,\n  LLAMA_PROMPT_GUARD_2_22M.name,\n  GPT_OSS_20B.name,\n  GPT_OSS_120B.name,\n  GPT_OSS_SAFEGUARD_20B.name,\n  KIMI_K2_INSTRUCT_0905.name,\n  QWEN3_32B.name,\n] as const\n\n/**\n * Union type of all supported Groq chat model names.\n */\nexport type GroqChatModels = (typeof GROQ_CHAT_MODELS)[number]\n\n/**\n * Type-only map from Groq chat model name to its supported input modalities.\n */\nexport type GroqModelInputModalitiesByName = {\n  [LLAMA_3_1_8B_INSTANT.name]: typeof LLAMA_3_1_8B_INSTANT.supports.input\n  [LLAMA_3_3_70B_VERSATILE.name]: typeof LLAMA_3_3_70B_VERSATILE.supports.input\n  [LLAMA_4_MAVERICK_17B_128E_INSTRUCT.name]: typeof LLAMA_4_MAVERICK_17B_128E_INSTRUCT.supports.input\n  [LLAMA_4_SCOUT_17B_16E_INSTRUCT.name]: typeof LLAMA_4_SCOUT_17B_16E_INSTRUCT.supports.input\n  [LLAMA_GUARD_4_12B.name]: typeof LLAMA_GUARD_4_12B.supports.input\n  [LLAMA_PROMPT_GUARD_2_86M.name]: typeof LLAMA_PROMPT_GUARD_2_86M.supports.input\n  [LLAMA_PROMPT_GUARD_2_22M.name]: typeof LLAMA_PROMPT_GUARD_2_22M.supports.input\n  [GPT_OSS_20B.name]: typeof GPT_OSS_20B.supports.input\n  [GPT_OSS_120B.name]: typeof GPT_OSS_120B.supports.input\n  [GPT_OSS_SAFEGUARD_20B.name]: typeof GPT_OSS_SAFEGUARD_20B.supports.input\n  [KIMI_K2_INSTRUCT_0905.name]: typeof KIMI_K2_INSTRUCT_0905.supports.input\n  [QWEN3_32B.name]: typeof QWEN3_32B.supports.input\n}\n\n/**\n * Type-only map from Groq chat model name to its provider options type.\n */\nexport type GroqChatModelProviderOptionsByName = {\n  [K in (typeof GROQ_CHAT_MODELS)[number]]: GroqTextProviderOptions\n}\n\n/**\n * Type-only map from Groq chat model name to its supported provider tools.\n * Groq exposes no provider-specific tool factories, so every model gets an\n * empty tuple. This ensures that passing an Anthropic/OpenAI ProviderTool to\n * a Groq adapter produces a compile-time type error.\n */\nexport type GroqChatModelToolCapabilitiesByName = {\n  [LLAMA_3_1_8B_INSTANT.name]: typeof LLAMA_3_1_8B_INSTANT.supports.tools\n  [LLAMA_3_3_70B_VERSATILE.name]: typeof LLAMA_3_3_70B_VERSATILE.supports.tools\n  [LLAMA_4_MAVERICK_17B_128E_INSTRUCT.name]: typeof LLAMA_4_MAVERICK_17B_128E_INSTRUCT.supports.tools\n  [LLAMA_4_SCOUT_17B_16E_INSTRUCT.name]: typeof LLAMA_4_SCOUT_17B_16E_INSTRUCT.supports.tools\n  [LLAMA_GUARD_4_12B.name]: typeof LLAMA_GUARD_4_12B.supports.tools\n  [LLAMA_PROMPT_GUARD_2_86M.name]: typeof LLAMA_PROMPT_GUARD_2_86M.supports.tools\n  [LLAMA_PROMPT_GUARD_2_22M.name]: typeof LLAMA_PROMPT_GUARD_2_22M.supports.tools\n  [GPT_OSS_20B.name]: typeof GPT_OSS_20B.supports.tools\n  [GPT_OSS_120B.name]: typeof GPT_OSS_120B.supports.tools\n  [GPT_OSS_SAFEGUARD_20B.name]: typeof GPT_OSS_SAFEGUARD_20B.supports.tools\n  [KIMI_K2_INSTRUCT_0905.name]: typeof KIMI_K2_INSTRUCT_0905.supports.tools\n  [QWEN3_32B.name]: typeof QWEN3_32B.supports.tools\n}\n\n/**\n * Type-only map from Groq TTS model name to its provider options type.\n */\nexport type GroqTTSModelProviderOptionsByName = {\n  [K in GroqTTSModel]: GroqTTSProviderOptions\n}\n\n/**\n * Resolves the provider options type for a specific Groq model.\n * Checks TTS models first, then chat models, then falls back to generic options.\n */\nexport type ResolveProviderOptions<TModel extends string> =\n  TModel extends GroqTTSModel\n    ? GroqTTSProviderOptions\n    : TModel extends keyof GroqChatModelProviderOptionsByName\n      ? GroqChatModelProviderOptionsByName[TModel]\n      : GroqTextProviderOptions\n\n/**\n * Resolve input modalities for a specific model.\n * If the model has explicit modalities in the map, use those; otherwise use text only.\n */\nexport type ResolveInputModalities<TModel extends string> =\n  TModel extends keyof GroqModelInputModalitiesByName\n    ? GroqModelInputModalitiesByName[TModel]\n    : readonly ['text']\n\n/**\n * All supported Groq transcription model identifiers.\n */\nexport const GROQ_TRANSCRIPTION_MODELS = [\n  'whisper-large-v3-turbo',\n  'whisper-large-v3',\n] as const\n\n/**\n * Union type of all supported Groq transcription model names.\n */\nexport type GroqTranscriptionModel = (typeof GROQ_TRANSCRIPTION_MODELS)[number]\n\n// ============================================================================\n// TTS Models\n// ============================================================================\n\nconst ORPHEUS_V1_ENGLISH = {\n  name: 'canopylabs/orpheus-v1-english',\n  pricing: {\n    input: {\n      normal: 22,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['audio'],\n    endpoints: ['tts'],\n    features: [],\n  },\n} as const satisfies ModelMeta<GroqTTSProviderOptions>\n\nconst ORPHEUS_ARABIC_SAUDI = {\n  name: 'canopylabs/orpheus-arabic-saudi',\n  pricing: {\n    input: {\n      normal: 40,\n    },\n  },\n  supports: {\n    input: ['text'],\n    output: ['audio'],\n    endpoints: ['tts'],\n    features: [],\n  },\n} as const satisfies ModelMeta<GroqTTSProviderOptions>\n\n/**\n * All supported Groq TTS model identifiers.\n */\nexport const GROQ_TTS_MODELS = [\n  ORPHEUS_V1_ENGLISH.name,\n  ORPHEUS_ARABIC_SAUDI.name,\n] as const\n\n/**\n * Union type of all supported Groq TTS model names.\n */\nexport type GroqTTSModel = (typeof GROQ_TTS_MODELS)[number]\n"],"names":[],"mappings":"AAsCA,MAAM,0BAA0B;AAAA,EAC9B,MAAM;AAkBR;AAEA,MAAM,qCAAqC;AAAA,EACzC,MAAM;AAkBR;AAEA,MAAM,iCAAiC;AAAA,EACrC,MAAM;AAkBR;AAEA,MAAM,oBAAoB;AAAA,EACxB,MAAM;AAkBR;AAEA,MAAM,2BAA2B;AAAA,EAC/B,MAAM;AAkBR;AAEA,MAAM,uBAAuB;AAAA,EAC3B,MAAM;AAkBR;AAEA,MAAM,2BAA2B;AAAA,EAC/B,MAAM;AAkBR;AAEA,MAAM,eAAe;AAAA,EACnB,MAAM;AA2BR;AAEA,MAAM,wBAAwB;AAAA,EAC5B,MAAM;AA4BR;AAEA,MAAM,cAAc;AAAA,EAClB,MAAM;AA2BR;AAEA,MAAM,wBAAwB;AAAA,EAC5B,MAAM;AAmBR;AAEA,MAAM,YAAY;AAAA,EAChB,MAAM;AAkBR;AAKO,MAAM,mBAAmB;AAAA,EAC9B,qBAAqB;AAAA,EACrB,wBAAwB;AAAA,EACxB,mCAAmC;AAAA,EACnC,+BAA+B;AAAA,EAC/B,kBAAkB;AAAA,EAClB,yBAAyB;AAAA,EACzB,yBAAyB;AAAA,EACzB,YAAY;AAAA,EACZ,aAAa;AAAA,EACb,sBAAsB;AAAA,EACtB,sBAAsB;AAAA,EACtB,UAAU;AACZ;AAmFO,MAAM,4BAA4B;AAAA,EACvC;AAAA,EACA;AACF;AAWA,MAAM,qBAAqB;AAAA,EACzB,MAAM;AAYR;AAEA,MAAM,uBAAuB;AAAA,EAC3B,MAAM;AAYR;AAKO,MAAM,kBAAkB;AAAA,EAC7B,mBAAmB;AAAA,EACnB,qBAAqB;AACvB;"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tanstack/ai-groq",
-  "version": "0.4.16",
+  "version": "0.5.1",
   "description": "Groq adapter for TanStack AI low-latency chat, tool calling, and structured outputs.",
   "author": "Tanner Linsley",
   "license": "MIT",
@@ -53,12 +53,12 @@
   },
   "peerDependencies": {
     "zod": "^4.0.0",
-    "@tanstack/ai": "^0.39.0"
+    "@tanstack/ai": "^0.40.0"
   },
   "dependencies": {
     "openai": "^6.41.0",
     "@tanstack/ai-utils": "0.3.1",
-    "@tanstack/openai-base": "0.9.6"
+    "@tanstack/openai-base": "0.9.7"
   },
   "scripts": {
     "build": "vite build",

package/src/adapters/transcription.ts

@@ -0,0 +1,334 @@
+import { BaseTranscriptionAdapter } from '@tanstack/ai/adapters'
+import { base64ToArrayBuffer, generateId } from '@tanstack/ai-utils'
+import { getGroqApiKeyFromEnv, withGroqDefaults } from '../utils/client'
+import type {
+  TranscriptionOptions,
+  TranscriptionResult,
+  TranscriptionSegment,
+} from '@tanstack/ai'
+import type { GroqTranscriptionModel } from '../model-meta'
+import type { GroqTranscriptionProviderOptions } from '../audio/transcription-provider-options'
+import type { GroqClientConfig } from '../utils/client'
+
+/**
+ * Configuration for the Groq Transcription adapter.
+ */
+export interface GroqTranscriptionConfig extends GroqClientConfig {}
+
+/**
+ * Flattens the `openai` SDK's `HeadersLike` config value into a plain record so
+ * it can be merged into the raw `fetch` request this adapter issues. Handles
+ * the shapes callers actually pass (`Headers`, an entries array, or a plain
+ * object); null/undefined values are dropped.
+ *
+ * ponytail: doesn't unwrap the SDK's internal `NullableHeaders` class; forward
+ * that shape here if the SDK ever hands it to adapter config.
+ */
+function normalizeHeaders(
+  headers: GroqTranscriptionConfig['defaultHeaders'],
+): Record<string, string> {
+  const out: Record<string, string> = {}
+  if (!headers) return out
+  const assign = (key: string, value: unknown) => {
+    if (value != null) out[key] = String(value)
+  }
+  if (headers instanceof Headers) {
+    headers.forEach((value, key) => assign(key, value))
+  } else if (Array.isArray(headers)) {
+    for (const [key, value] of headers) assign(key, value)
+  } else {
+    for (const [key, value] of Object.entries(headers)) assign(key, value)
+  }
+  return out
+}
+
+// Shape of Groq's verbose_json transcription response
+interface GroqVerboseTranscriptionResponse {
+  task?: string
+  language?: string
+  duration?: number
+  text: string
+  segments?: Array<{
+    id: number
+    seek?: number
+    start: number
+    end: number
+    text: string
+    tokens?: Array<number>
+    temperature?: number
+    avg_logprob: number
+    compression_ratio?: number
+    no_speech_prob?: number
+  }>
+  words?: Array<{ word: string; start: number; end: number }>
+  x_groq?: { id?: string }
+}
+
+// Shape of Groq's json transcription response
+interface GroqJsonTranscriptionResponse {
+  text: string
+  x_groq?: { id?: string }
+}
+
+/**
+ * Groq Transcription (Speech-to-Text) Adapter
+ *
+ * Tree-shakeable adapter for Groq audio transcription. Supports
+ * whisper-large-v3 and whisper-large-v3-turbo.
+ *
+ * Features:
+ * - Audio file uploads (File, Blob, ArrayBuffer, base64/data URL)
+ * - Remote audio URLs passed directly via Groq's `url` field — no upload needed
+ * - Verbose JSON response with segment and word timestamps
+ * - Language detection or specification (ISO-639-1)
+ * - Confidence scores derived from segment avg_logprob
+ */
+export class GroqTranscriptionAdapter<
+  TModel extends GroqTranscriptionModel,
+> extends BaseTranscriptionAdapter<TModel, GroqTranscriptionProviderOptions> {
+  readonly name = 'groq' as const
+
+  private readonly apiKey: string
+  private readonly baseURL: string
+  private readonly defaultHeaders: Record<string, string>
+
+  constructor(config: GroqTranscriptionConfig, model: TModel) {
+    super(model, {})
+    const resolved = withGroqDefaults(config)
+    this.apiKey = resolved.apiKey
+    this.baseURL = resolved.baseURL ?? 'https://api.groq.com/openai/v1'
+    this.defaultHeaders = normalizeHeaders(resolved.defaultHeaders)
+  }
+
+  async transcribe(
+    options: TranscriptionOptions<GroqTranscriptionProviderOptions>,
+  ): Promise<TranscriptionResult> {
+    const { model, audio, language, prompt, responseFormat, modelOptions } =
+      options
+
+    // Groq's transcription endpoint only accepts 'json', 'text', and
+    // 'verbose_json'. Reject 'srt'/'vtt' up front so callers get a clear
+    // message instead of an opaque Groq HTTP error.
+    if (responseFormat === 'srt' || responseFormat === 'vtt') {
+      throw new Error(
+        `Groq transcription does not support responseFormat='${responseFormat}'. ` +
+          `Supported values: 'json', 'text', 'verbose_json'.`,
+      )
+    }
+
+    // Default to verbose_json so callers get language, duration, and timestamps
+    // without having to opt in explicitly. Both Groq whisper models support it.
+    const effectiveFormat = responseFormat ?? 'verbose_json'
+    const useVerbose = effectiveFormat === 'verbose_json'
+
+    const form = new FormData()
+    form.append('model', model)
+    form.append('response_format', effectiveFormat)
+    if (language !== undefined) form.append('language', language)
+    if (prompt !== undefined) form.append('prompt', prompt)
+    if (modelOptions?.temperature !== undefined) {
+      form.append('temperature', String(modelOptions.temperature))
+    }
+    if (modelOptions?.timestamp_granularities !== undefined) {
+      for (const g of modelOptions.timestamp_granularities) {
+        form.append('timestamp_granularities[]', g)
+      }
+    }
+
+    // HTTP/HTTPS URLs are forwarded directly via Groq's `url` field, which
+    // avoids a round-trip upload. All other inputs (File, Blob, ArrayBuffer,
+    // base64, data URL) are converted to a File and sent as `file`.
+    if (typeof audio === 'string' && /^https?:\/\//.test(audio)) {
+      form.append('url', audio)
+    } else {
+      form.append('file', this.prepareAudioFile(audio))
+    }
+
+    try {
+      options.logger.request(
+        `activity=transcription provider=${this.name} model=${model} verbose=${useVerbose}`,
+        { provider: this.name, model },
+      )
+
+      const response = await fetch(`${this.baseURL}/audio/transcriptions`, {
+        method: 'POST',
+        headers: {
+          ...this.defaultHeaders,
+          Authorization: `Bearer ${this.apiKey}`,
+        },
+        body: form,
+      })
+
+      if (!response.ok) {
+        const body = await response
+          .json()
+          .catch(() => null as Record<string, unknown> | null)
+        const message =
+          (body?.error as { message?: string } | undefined)?.message ??
+          `Groq API error ${response.status}`
+        throw new Error(message)
+      }
+
+      if (useVerbose) {
+        const data = (await response.json()) as GroqVerboseTranscriptionResponse
+        const requestId = data.x_groq?.id ?? generateId(this.name)
+
+        // `TranscriptionResult` declares optional fields without `| undefined`,
+        // so under exactOptionalPropertyTypes we must omit absent fields rather
+        // than assigning `undefined`.
+        const segments = data.segments?.map(
+          (seg): TranscriptionSegment => ({
+            id: seg.id,
+            start: seg.start,
+            end: seg.end,
+            text: seg.text,
+            confidence: Math.exp(seg.avg_logprob),
+          }),
+        )
+        const words = data.words?.map((w) => ({
+          word: w.word,
+          start: w.start,
+          end: w.end,
+        }))
+
+        return {
+          id: requestId,
+          model,
+          text: data.text,
+          ...(data.language !== undefined && { language: data.language }),
+          ...(data.duration !== undefined && { duration: data.duration }),
+          ...(segments !== undefined && { segments }),
+          ...(words !== undefined && { words }),
+        }
+      } else if (effectiveFormat === 'text') {
+        const text = await response.text()
+        return {
+          id: generateId(this.name),
+          model,
+          text,
+          ...(language !== undefined && { language }),
+        }
+      } else {
+        const data = (await response.json()) as GroqJsonTranscriptionResponse
+        return {
+          id: data.x_groq?.id ?? generateId(this.name),
+          model,
+          text: data.text,
+          ...(language !== undefined && { language }),
+        }
+      }
+    } catch (error: unknown) {
+      options.logger.errors(`${this.name}.transcribe fatal`, {
+        error,
+        source: `${this.name}.transcribe`,
+      })
+      throw error
+    }
+  }
+
+  private prepareAudioFile(audio: string | File | Blob | ArrayBuffer): File {
+    if (typeof File !== 'undefined' && audio instanceof File) {
+      return audio
+    }
+    if (typeof Blob !== 'undefined' && audio instanceof Blob) {
+      this.ensureFileSupport()
+      return new File([audio], 'audio.mp3', {
+        type: audio.type || 'audio/mpeg',
+      })
+    }
+    if (typeof ArrayBuffer !== 'undefined' && audio instanceof ArrayBuffer) {
+      this.ensureFileSupport()
+      return new File([audio], 'audio.mp3', { type: 'audio/mpeg' })
+    }
+    if (typeof audio === 'string') {
+      this.ensureFileSupport()
+
+      if (audio.startsWith('data:')) {
+        const parts = audio.split(',')
+        const header = parts[0]
+        const base64Data = parts[1] || ''
+        const mimeMatch = header?.match(/data:([^;]+)/)
+        const mimeType = mimeMatch?.[1] || 'audio/mpeg'
+        const bytes = base64ToArrayBuffer(base64Data)
+        const extension = mimeType.split('/')[1] || 'mp3'
+        return new File([bytes], `audio.${extension}`, { type: mimeType })
+      }
+
+      const bytes = base64ToArrayBuffer(audio)
+      return new File([bytes], 'audio.mp3', { type: 'audio/mpeg' })
+    }
+
+    throw new Error('Invalid audio input type')
+  }
+
+  // Throws on Node < 20 where the global `File` constructor is unavailable.
+  private ensureFileSupport(): void {
+    if (typeof File === 'undefined') {
+      throw new Error(
+        '`File` is not available in this environment. ' +
+          'Use Node.js 20 or newer, or pass a File object directly.',
+      )
+    }
+  }
+}
+
+/**
+ * Creates a Groq transcription adapter with an explicit API key.
+ * Type resolution happens here at the call site.
+ *
+ * @param model - The model name (e.g., 'whisper-large-v3-turbo')
+ * @param apiKey - Your Groq API key
+ * @param config - Optional additional configuration
+ * @returns Configured Groq transcription adapter instance
+ *
+ * @example
+ * ```typescript
+ * const adapter = createGroqTranscription('whisper-large-v3-turbo', 'gsk_...');
+ *
+ * const result = await generateTranscription({
+ *   adapter,
+ *   audio: audioFile,
+ *   language: 'en',
+ * });
+ * ```
+ */
+export function createGroqTranscription<TModel extends GroqTranscriptionModel>(
+  model: TModel,
+  apiKey: string,
+  config?: Omit<GroqTranscriptionConfig, 'apiKey'>,
+): GroqTranscriptionAdapter<TModel> {
+  return new GroqTranscriptionAdapter({ apiKey, ...config }, model)
+}
+
+/**
+ * Creates a Groq transcription adapter using the `GROQ_API_KEY` environment
+ * variable. Type resolution happens here at the call site.
+ *
+ * Looks for `GROQ_API_KEY` in:
+ * - `process.env` (Node.js)
+ * - `window.env` (browser with injected env)
+ *
+ * @param model - The model name (e.g., 'whisper-large-v3-turbo')
+ * @param config - Optional configuration (excluding apiKey which is auto-detected)
+ * @returns Configured Groq transcription adapter instance
+ * @throws Error if GROQ_API_KEY is not found in environment
+ *
+ * @example
+ * ```typescript
+ * const adapter = groqTranscription('whisper-large-v3-turbo');
+ *
+ * const result = await generateTranscription({
+ *   adapter,
+ *   audio: 'https://example.com/audio.mp3',
+ * });
+ *
+ * console.log(result.text)
+ * ```
+ */
+export function groqTranscription<TModel extends GroqTranscriptionModel>(
+  model: TModel,
+  config?: Omit<GroqTranscriptionConfig, 'apiKey'>,
+): GroqTranscriptionAdapter<TModel> {
+  const apiKey = getGroqApiKeyFromEnv()
+  return createGroqTranscription(model, apiKey, config)
+}

package/src/adapters/tts.ts

@@ -0,0 +1,168 @@
+import OpenAI from 'openai'
+import { BaseTTSAdapter } from '@tanstack/ai/adapters'
+import { toRunErrorPayload } from '@tanstack/ai/adapter-internals'
+import { arrayBufferToBase64, generateId } from '@tanstack/ai-utils'
+import { getGroqApiKeyFromEnv, withGroqDefaults } from '../utils/client'
+import { validateAudioInput } from '../audio/audio-provider-options'
+import type { TTSOptions, TTSResult } from '@tanstack/ai'
+import type OpenAI_SDK from 'openai'
+import type { GroqTTSModel } from '../model-meta'
+import type { GroqTTSProviderOptions } from '../audio/tts-provider-options'
+import type { GroqClientConfig } from '../utils'
+
+/**
+ * Configuration for Groq TTS adapter
+ */
+export interface GroqTTSConfig extends GroqClientConfig {}
+
+/**
+ * Groq Text-to-Speech Adapter
+ *
+ * Tree-shakeable adapter for Groq TTS functionality. Groq exposes an
+ * OpenAI-compatible `/audio/speech` endpoint, so the adapter drives it with
+ * the OpenAI SDK via a `baseURL` override (the same pattern as the Groq text
+ * adapter).
+ *
+ * Supports `canopylabs/orpheus-v1-english` and
+ * `canopylabs/orpheus-arabic-saudi`.
+ *
+ * Features:
+ * - English voices: autumn(f), diana(f), hannah(f), austin(m), daniel(m), troy(m)
+ * - Arabic voices: fahad(m), sultan(m), lulwa(f), noura(f)
+ * - Output formats: flac, mp3, mulaw, ogg, wav (default wav)
+ * - Speed control
+ * - Configurable sample rate via `modelOptions`
+ */
+export class GroqTTSAdapter<TModel extends GroqTTSModel> extends BaseTTSAdapter<
+  TModel,
+  GroqTTSProviderOptions
+> {
+  readonly name = 'groq' as const
+
+  protected client: OpenAI
+
+  constructor(config: GroqTTSConfig, model: TModel) {
+    super(model, {})
+    this.client = new OpenAI(withGroqDefaults(config))
+  }
+
+  async generateSpeech(
+    options: TTSOptions<GroqTTSProviderOptions>,
+  ): Promise<TTSResult> {
+    const { model, text, voice, format, speed, modelOptions } = options
+
+    validateAudioInput({ input: text, model: this.model })
+
+    // Spreading optional inputs conditionally keeps the request compatible
+    // with the vendor SDK shape under exactOptionalPropertyTypes. `sample_rate`
+    // is a Groq-only body field carried via modelOptions.
+    const request: OpenAI_SDK.Audio.SpeechCreateParams = {
+      model,
+      input: text,
+      voice: voice ?? 'autumn',
+      response_format: format ?? 'wav',
+      ...(speed !== undefined && { speed }),
+      ...(modelOptions ?? {}),
+    }
+
+    try {
+      options.logger.request(
+        `activity=tts provider=${this.name} model=${model} format=${request.response_format ?? 'default'} voice=${request.voice}`,
+        { provider: this.name, model },
+      )
+      const response = await this.client.audio.speech.create(request)
+
+      const arrayBuffer = await response.arrayBuffer()
+      const base64 = arrayBufferToBase64(arrayBuffer)
+
+      const outputFormat = request.response_format ?? 'wav'
+      const contentType = this.getContentType(outputFormat)
+
+      return {
+        id: generateId(this.name),
+        model,
+        audio: base64,
+        format: outputFormat,
+        contentType,
+      }
+    } catch (error: unknown) {
+      // Narrow before logging: raw SDK errors can carry request metadata
+      // (including auth headers) which we must never surface to user loggers.
+      options.logger.errors(`${this.name}.generateSpeech fatal`, {
+        error: toRunErrorPayload(error, `${this.name}.generateSpeech failed`),
+        source: `${this.name}.generateSpeech`,
+      })
+      throw error
+    }
+  }
+
+  private getContentType(format: string): string {
+    const contentTypes: Record<string, string> = {
+      flac: 'audio/flac',
+      mp3: 'audio/mpeg',
+      mulaw: 'audio/basic',
+      ogg: 'audio/ogg',
+      wav: 'audio/wav',
+    }
+    return contentTypes[format] || 'audio/wav'
+  }
+}
+
+/**
+ * Creates a Groq speech adapter with explicit API key.
+ * Type resolution happens here at the call site.
+ *
+ * @param model - The model name (e.g., 'canopylabs/orpheus-v1-english')
+ * @param apiKey - Your Groq API key
+ * @param config - Optional additional configuration
+ * @returns Configured Groq speech adapter instance with resolved types
+ *
+ * @example
+ * ```typescript
+ * const adapter = createGroqSpeech('canopylabs/orpheus-v1-english', 'gsk_...')
+ *
+ * const result = await generateSpeech({
+ *   adapter,
+ *   text: 'Hello, world!',
+ *   voice: 'autumn',
+ * })
+ * ```
+ */
+export function createGroqSpeech<TModel extends GroqTTSModel>(
+  model: TModel,
+  apiKey: string,
+  config?: Omit<GroqTTSConfig, 'apiKey'>,
+): GroqTTSAdapter<TModel> {
+  return new GroqTTSAdapter({ apiKey, ...config }, model)
+}
+
+/**
+ * Creates a Groq speech adapter with automatic API key detection from
+ * environment variables.
+ *
+ * Looks for `GROQ_API_KEY` in the environment.
+ *
+ * @param model - The model name (e.g., 'canopylabs/orpheus-v1-english')
+ * @param config - Optional configuration (excluding apiKey which is auto-detected)
+ * @returns Configured Groq speech adapter instance with resolved types
+ * @throws Error if GROQ_API_KEY is not found in environment
+ *
+ * @example
+ * ```typescript
+ * const adapter = groqSpeech('canopylabs/orpheus-v1-english')
+ *
+ * const result = await generateSpeech({
+ *   adapter,
+ *   text: 'Welcome to TanStack AI!',
+ *   voice: 'autumn',
+ *   format: 'wav',
+ * })
+ * ```
+ */
+export function groqSpeech<TModel extends GroqTTSModel>(
+  model: TModel,
+  config?: Omit<GroqTTSConfig, 'apiKey'>,
+): GroqTTSAdapter<TModel> {
+  const apiKey = getGroqApiKeyFromEnv()
+  return createGroqSpeech(model, apiKey, config)
+}

package/src/audio/audio-provider-options.ts

@@ -0,0 +1,25 @@
+/**
+ * Common audio provider options for Groq audio endpoints.
+ */
+export interface AudioProviderOptions {
+  /**
+   * The text to generate audio for.
+   * Maximum length is 200 characters.
+   * Use [directions] for vocal control (English voices only).
+   */
+  input: string
+  /**
+   * The audio model to use for generation.
+   */
+  model: string
+}
+
+/**
+ * Validates that the audio input text does not exceed the maximum length.
+ * @throws Error if input text exceeds 200 characters
+ */
+export const validateAudioInput = (options: AudioProviderOptions) => {
+  if (options.input.length > 200) {
+    throw new Error('Input text exceeds maximum length of 200 characters.')
+  }
+}

package/src/audio/transcription-provider-options.ts

@@ -0,0 +1,20 @@
+/**
+ * Groq-specific options for audio transcription.
+ *
+ * These fields extend the shared `TranscriptionOptions` and are forwarded
+ * verbatim to the Groq transcription endpoint.
+ */
+export interface GroqTranscriptionProviderOptions {
+  /**
+   * Sampling temperature between 0 and 1. Lower values produce more
+   * deterministic output. Groq recommends 0 (the default) for most use cases.
+   */
+  temperature?: number
+
+  /**
+   * Granularity levels to include when `response_format` is `verbose_json`.
+   * Pass `['word']`, `['segment']`, or both to control which timestamp arrays
+   * appear in the result.
+   */
+  timestamp_granularities?: Array<'word' | 'segment'>
+}