> For clean Markdown of any page, append `.md` to the page URL. > For a complete documentation index, see https://docs.sarvam.ai/llms.txt. > For full documentation content in one file, see https://docs.sarvam.ai/llms-full.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.sarvam.ai/_mcp/server. # Change Log > Stay updated with the latest Sarvam AI API changes, new features, and improvements. Track releases for speech-to-text, translation, chat completion, and other services. # August 2026 ### Speech to Text (STT) **Realtime API \[Model APIs]**: Introduced a new [Realtime API](/api/api-guides-tutorials/speech-to-text/realtime-streaming) - `saaras:v3-realtime` over WebSocket - purpose-built for voice agents and live transcription. It adds true partial (interim) transcripts as the user speaks, live mid-stream reconfiguration, and millisecond-based VAD tuning. This supersedes the legacy streaming WebSocket for new voice-agent and live-transcription work; the older `saaras:v3` streaming WebSocket remains generally available and is now documented under Legacy. ### Open-Source Models **GLM-5.2 and Gemma 4 31B (Beta) \[Model APIs]**: Introduced two open-source chat models served on Sarvam infrastructure via the new `/v2/chat/completions` endpoint. [GLM-5.2](/api/getting-started/models/open-source/glm-5-2) (`glm5.2`) offers a 512K-token context window with tool calling and visible reasoning, for tasks that exceed Sarvam's own 128K context. [Gemma 4 31B](/api/getting-started/models/open-source/gemma-4-31b) (`gemma4`) adds image input alongside tool calling. Neither model is tuned for Indian languages - for Indian-language workloads, continue using [Sarvam-105B](/api/getting-started/models/sarvam-105b) or [Sarvam Vision](/api/getting-started/models/sarvam-vision). Beta access is granted per API key; [contact us](/api/getting-started/help) to request whitelisting. ### Self-Hosted Deployments **Self-Hosted Deployments \[Model APIs]**: Sarvam models can now run inside your own AWS account via [Self-Hosted Deployments](/api/self-hosted/introduction). Subscribe to a model package on the AWS Marketplace and deploy Saaras v3, Bulbul v3, or Sarvam Vision as an Amazon SageMaker endpoint in your own VPC - same models as the Managed API, but your audio and documents never leave your infrastructure. ### Pricing **LLM Pricing Updated \[Model APIs]**: `sarvam-105b` pricing increased to ₹29.28 / ₹10.98 / ₹73.2 per 1M tokens (input / cached input / output), up from ₹4 / ₹2.5 / ₹16, effective August 5, 2026. `sarvam-105b-conversations` is priced the same as `sarvam-105b`. Pricing was also published for the beta open-source models - [GLM-5.2](/api/getting-started/models/open-source/glm-5-2) at ₹128.1 / ₹23.79 / ₹402.6 and [Gemma 4 31B](/api/getting-started/models/open-source/gemma-4-31b) at ₹36.6 / ₹13.73 / ₹91.5 - with GLM-5.2's reasoning tokens billed as output. See the [pricing page](/api/getting-started/pricing) for details. ### Voice Agents **Renamed from Conversations \[Voice Agents]**: The product formerly called Conversations (internally, "Samvaad") is now [Voice Agents](/conversations/overview), an end-to-end stack for building, deploying, and monitoring voice agents on telephony. References to `platform.sarvam.ai` were updated accordingly. **Code Tools \[Voice Agents]**: Added [Code Tools](/conversations/build/tools/code-tools), letting you upload a custom Python file to define tools that go beyond the built-in API and data-validation tools (enterprise feature). ### Content Agents **Dubbing Restructured \[Content Agents]**: Reorganized [Dubbing](/creative-dubbing) docs into a hub with dedicated Voices and Job Lifecycle pages, and documented `editor_flow` pricing: ₹40 / ₹38 / ₹36 per minute (Starter / Pro / Enterprise) with `editor_flow: false`, and ₹80 / ₹75 / ₹72 per minute - 2× the base rate - with `editor_flow: true`. ### Integrations **Vobiz \[Model APIs]**: Added a Vobiz telephony integration guide for [Voice Agents](/conversations/overview). --- # July 2026 ### Chat Completion **Sarvam-30B Deprecated**: `sarvam-30b` has been fully deprecated and moved to the Deprecated Models section, and is no longer available through the Chat Completions API. [Sarvam-105B](/api/getting-started/models/sarvam-105b) is now the only supported Sarvam chat model - migrate any remaining `sarvam-30b` integrations to `sarvam-105b`. ### Speech to Text (STT) **Saaras v4 Added**: Introduced [`saaras:v4`](/api/getting-started/models/saaras), available alongside `saaras:v3` across the REST, WebSocket, and Batch Speech-to-Text endpoints. It supports the same five output modes (`transcribe`, `translate`, `verbatim`, `translit`, `codemix`) and adds Global English alongside Indian English, on top of the existing 22 Indic languages. `saaras:v3` remains the default, recommended model. ### Document AI **Document AI Launched**: Replaced the Document Digitization API with [Document AI](/api/api-guides-tutorials/document-intelligence/overview) (`/doc-ai/v1`), powered by an upgraded Sarvam Vision 1.5 trained for both OCR and key-value extraction. Two endpoints: **Digitise** for full-document OCR with layout and tables preserved, and **Extract** for schema-based field extraction to JSON, CSV, or XLSX. The earlier Document Digitization API is now legacy. ### Content Agents **Dubbing API Launched (Beta)**: Added job endpoints for the [Content Agents](/creative/studio-overview) [Dubbing](/creative-dubbing) feature, which localizes video or audio into Indian languages while preserving the original speaker's voice identity and delivery. ### Products **Conversations and Work Agents Launched**: Opened up two new products. Conversations (renamed [Voice Agents](/conversations/overview) the following month) is an end-to-end stack for building, deploying, and monitoring voice agents on telephony, previously run only internally. [Work Agents](/work-agents/overview) is an agentic workspace for ad-hoc tasks and scheduled agents, with Connectors into tools like Gmail, Slack, GitHub, and Notion. Also renamed Sarvam Pages to [Doc Agents](/docai/getting-started/overview) and Sarvam Creative to [Content Agents](/creative/studio-overview), dropping the "Sarvam" prefix across the product lineup. ### Integrations **Telephony, Frameworks & Migrations**: Added integration guides for Twilio, Plivo, and Exotel telephony, plus Vapi and LangChain; LiveKit and Pipecat production-tuning guides for [Voice Agents](/conversations/overview); and migration guides for developers switching from Cartesia, Deepgram, ElevenLabs, or Gemini. --- # June 2026 ### Chat Completion **Sarvam-M Deprecated**: `sarvam-m` has been deprecated and is no longer accepted by the Chat Completions API — requests that pass `model="sarvam-m"` will fail. The supported chat models are now `sarvam-30b` (64K context) and `sarvam-105b` (128K context). Existing integrations should migrate to [Sarvam-30B](/api/getting-started/models/sarvam-30b) or [Sarvam-105B](/api/getting-started/models/sarvam-105b) for better quality and longer context. ### Speech to Text (STT) **Default Model is now `saaras:v3`**: The `/speech-to-text` REST and Batch endpoints now default to `saaras:v3` when no `model` is specified (previously `saarika:v2.5`). Requests that omit `model` will use the new model and respect the `mode` parameter (`transcribe`, `translate`, `verbatim`, `translit`, `codemix`). To keep the previous behavior, pass `model="saarika:v2.5"` explicitly. `saarika:v2.5` remains available as a legacy model. **Translate via `/speech-to-text`**: The dedicated `/speech-to-text-translate` endpoint is now legacy (`saaras:v2.5`). For state-of-the-art translation, use the `/speech-to-text` endpoint with `model="saaras:v3"` and `mode="translate"`. --- # May 2026 ### Chat Completion **Removed Model Variants**: Retired the fixed-context `sarvam-30b-16k` and `sarvam-105b-32k` variants. The base `sarvam-30b` and `sarvam-105b` models now serve their full context windows (64K and 128K tokens), so you no longer need to pick a context-specific variant. ### Pricing **Document Digitization**: Document Intelligence has been renamed to **Document Digitization** and repriced to ₹0.5/page, down from ₹1.5/page, making bulk document processing significantly cheaper. **Free Credits**: New accounts now receive ₹100 in free credits (previously ₹1,000) — worth noting if you provision trial accounts programmatically. --- # April 2026 ### Chat Completion **Updated Parameter Defaults**: Adjusted the default values for `reasoning_effort` and `max_tokens` in the Chat Completions API. If you rely on the defaults rather than setting these explicitly, response length and latency may change — set them explicitly to lock in previous behavior. **Deprecated Model Variants**: Marked the `sarvam-105b-32k` and `sarvam-30b-16k` variants as deprecated ahead of their removal, signalling customers to move to the base `sarvam-30b` / `sarvam-105b` models. ### Speech to Text (STT) **VAD Parameters**: Added Voice Activity Detection parameters to STT WebSocket streaming, giving you control over how speech segments are detected and finalized in real-time transcription. ### Integrations **n8n**: Published the Sarvam AI n8n integration, letting you call Sarvam APIs from n8n workflows as drop-in automation nodes. --- # March 2026 ### Chat Completion **Removed Gemma Models**: Removed the Gemma model IDs from the `SarvamModelIds` enum. These are no longer accepted values for the `model` parameter on the Chat Completions API. **Removed Wiki Grounding**: Wiki grounding is no longer supported for `sarvam-30b` and `sarvam-105b`. Requests that previously enabled grounding on these models should drop the parameter. **SDK Chat Support**: Added Chat Completions (LLM) support to the Python and TypeScript SDKs, so you can call the chat API through a typed client instead of raw HTTP requests. ### Speech to Text (STT) **Pronunciation Dictionary v2**: Added pronunciation dictionary v2, giving you finer control over how specific words and names are transcribed. ### Document Digitization **Page Limit & Pricing**: Introduced a 10-page-per-request limit and updated pricing to ₹1.5/page. Split larger documents into 10-page batches to stay within the limit. --- # February 2026 ### Speech to Text (STT) **Saaras v3**: Migrated STT to `saaras:v3` and introduced a new `mode` parameter to control transcription behavior. v3 improves accuracy and expands language coverage; ASR v3 is also now available through the SDKs. ### Text to Speech (TTS) **Bulbul v3**: Promoted the Bulbul TTS model from `bulbul:v3-beta` to the stable `bulbul:v3`. Update your `model` value from the beta tag to `bulbul:v3`, which is now production-ready. ### Document Digitization **SDK Support**: Added Document Digitization support to the Python and JavaScript SDKs, so document parsing can be called through the official clients. --- # January 2026 ### Speech to Text (STT) **Removed `input_audio_codec`**: Removed the `input_audio_codec` parameter — the API now auto-detects the codec for supported formats. You can stop sending this parameter; it will be ignored. ### TTS & ASR **v3 Support**: Added TTS v3 and ASR v3 support across the SDKs, bringing the latest Bulbul and Saaras model generations to the official clients. --- # December 2025 ### Speech to Text (STT) **Added `input_audio_codec`**: Introduced the `input_audio_codec` parameter, letting you explicitly declare the audio codec when automatic detection isn't sufficient (for example, raw PCM streams). ### Translate & Transliterate **Deprecated `enable_preprocessing`**: Deprecated the `enable_preprocessing` parameter on the Translate and Transliterate APIs. Preprocessing behavior is now handled internally, so the parameter no longer needs to be set. ### SDKs **TypeScript SDK**: Released an updated TypeScript SDK version with improved generator settings and type coverage. --- # November 2025 ### Speech to Text (STT) **Sample Rate Configuration**: Added sample-rate configuration parameters to the STT API, so you can match the API to your source audio (e.g. 8 kHz telephony or 16 kHz) for more accurate transcription. --- # October 2025 ### Dashboard **API Key Usage Tracking**: You can now track the usage of each API key directly from your dashboard, making it easier to monitor and manage your API consumption. --- # September 2025 ### Speech to Text (STT) **Flush Signal in WebSocket**: STT and STT Translate WebSocket now support flush signal to finalize transcriptions cleanly between segments, enabling better control over transcription boundaries. **8 kHz Sample Rate Support**: Streaming STT now supports 8 kHz sample rate, making it easier to work with telephony and low-bandwidth audio applications. ### Text to Speech (TTS) **End Signal in WebSocket**: Sarvam TTS WebSocket now supports an end signal for smoother control of audio streams, allowing better integration with real-time applications. ### Dashboard **GST Invoices**: GST invoices are now available in the dashboard for streamlined billing and compliance. ### Pricing **Flexible Pricing Plans**: Released flexible pricing tiers to match different use cases, from prototyping to production. View details at [dashboard.sarvam.ai/pricing](https://dashboard.sarvam.ai/pricing). --- # August 2025 ### Speech to Text (STT) **PCM Audio Format Support**: Added support for PCM formats including `pcm_s16le`, `pcm_l16`, and `pcm_raw` for both STT and STT Translate APIs. * For most audio formats, our API automatically detects the codec * When using PCM formats, you must explicitly specify the `input_audio_codec` parameter * PCM files are only supported at 16kHz sample rate --- # July 2025 ### SDKs **Python SDK Long Audio Support**: Our Python SDK now supports processing long audio files up to 1 hour in duration through both synchronous and asynchronous methods. Ideal for processing: * Meetings, interviews, and call center recordings * Large-scale content processing pipelines Key Features: * Support for files up to 1 hour long * Speaker diarization and chunk-level timestamp support ### Speech to Text (STT) **STT WebSocket Update**: The Start Event and End Event is now returned by STT/STTT WebSockets, enabling better control and event handling in real-time transcription. --- # June 2025 ### Translation **Sarvam-Translate Launched**: Released `sarvam-translate:v1`, an open-weights model supporting 22 Indian languages. **Text Quickstart Updated**: Added `sarvam-translate` usage examples to the Python SDK Quickstart and Playground. ### Speech to Text (STT) **Real-Time STT via WebSocket**: Added WebSocket support for live transcription with ultra-low latency in both Python and JavaScript SDKs. **Audio Format Support Expanded**: STT now accepts `mp3`, `wav`, `aac`, `aiff`, `ogg/opus`, `flac`, `mp4/m4a`, and `amr` input formats. **Batch STT (Alpha)**: Introduced alpha support for batch transcription in Python SDK. Install via `pip install sarvamai==0.1.11a2`. **Saaras & Saarika WebSocket Support**: Both `saaras` and `saarika` models now support real-time streaming via WebSocket. ### Text to Speech (TTS) **Real-Time TTS Streaming**: Generate speech on the fly using WebSocket streaming. Available in Python and JavaScript SDKs. **Audio Format Support Expanded**: TTS now outputs in `mp3`, `linear16`, `mulaw`, `alaw`, `opus`, `flac`, `aac`, and `wav`. ### SDKs **Python SDK v4.23.2**: Includes real-time streaming support, batch transcription (alpha), new translation APIs, and model updates. **JavaScript SDK Updated**: Added real-time STT & TTS WebSocket support and updated documentation with streaming quickstarts. ### Models **New Model Added**: Introduced `sarvam-m` to the model family. **Model Updates**: `saarika` and `saaras` updated to `v2.5` with character limit improvements and stability enhancements. ### Documentation **Cookbooks Expanded**: Added examples for `sarvam-translate`, updated LID and chat completion cookbooks, updated with the SDK package and refreshed starter notebooks. **AI-Powered Docs Assistant**: Added an AI assistant to the documentation search bar for instant Q\&A and developer support. ### Dashboard **Usage Analytics Dashboard**: Released real-time API usage and credit tracking at [dashboard.sarvam.ai/usage](https://dashboard.sarvam.ai/usage). --- # May 2025 ### Models #### Sarvam-M Released Introduced `sarvam-m`, a 24B open-weights hybrid model based on Mistral Small. * Available now via [API Playground](https://dashboard.sarvam.ai/playground) and detailed in the [technical blog](https://sarvam.ai/blogs/sarvam-m). * Benchmarks added to the official documentation. #### Saarika v2.5 Released Improved voice quality and naturalness in `saarika:v2.5`, now available for use in the STT API. #### TTS WebSocket (Beta) Real-time streaming TTS now supported via WebSocket for beta users. Contact support to request access. ### API & Dashboard #### Upgraded Developer Dashboard * Unified dashboard with no-code playground and API key management in one place. * Easily test endpoints like LLM, TTS, STT, and Translate without writing code. * Prebuilt examples include Resume Translate and Hinglish Code Debug. #### API Playground Enhancements * Added full support for live testing of Sarvam-M, Sarvam-Translate, and TTS/STT APIs. * Playground features instant feedback and parameter tuning without leaving the dashboard. ### SDKs #### Official SDKs Released * Python: `pip install sarvamai` * JavaScript: `npm install sarvamai` * Abstracts away HTTP and response parsing with clean, unified methods across APIs. * Reduces integration time from hours to minutes. ### Documentation #### Revamped Developer Documentation * New cookbooks added with real-world SDK use cases including chat completion, translation, and speech. * Improved API navigation and content structure for faster discovery. * Code snippets updated for clarity and consistency with latest SDKs. #### Navigation Updates * Exposed core API endpoints directly in sidebar navigation. * Streamlined structure to help developers reach reference and guides faster. --- # April 2025 ### Text to Speech (TTS) #### Bulbul v2 Released Introduced `bulbul:v2`, the latest version of our Indian Text-to-Speech model. * More natural and expressive voice output with better emotional tone * Enhanced preprocessing for improved handling of mixed-language inputs #### Bulbul v1 Deprecation Notice `bulbul:v1` will be officially deprecated on April 30, 2025. Users should migrate to `bulbul:v2` to ensure uninterrupted service. #### 24kHz Audio Support Speech generation now supports `24kHz` sample rate for higher quality output. ### Speech to Text (STT) #### Batch ASR API Released New Batch ASR API allows uploading up to 20 audio files (up to 60 minutes each). * Ideal for calls, meetings, and long-form media * Available in both `speech-to-text` and `speech-to-text-translate` endpoints #### Real-Time API Improvements Improved transcription speed for real-time requests. * 3× faster processing for up to 30s audio snippets * Optimized for use cases like voice bots and instant assistants #### Streaming WebSocket (Beta) WebSocket-based real-time transcription now in beta. Early access available via the Sarvam Discord community. ### Text #### Language Detection Enhancements Added `source_language_code` to API responses. * Automatically detects input language * Improves performance on multilingual and code-switched text #### Indic Translation Upgrades Improved support for two-way translation between Indic and English. * Supports colloquial, modern, classical, and formal registers * Ideal for education, localization, and content creation platforms #### New APIs Introduced * **Transliterate API**\ Converts text between writing systems while preserving pronunciation * **Language Identification (LID) API**\ Detects both language and script from input ### Documentation #### Cookbook Updated * Added new real-world examples for Batch ASR and Bulbul v2 usage * Included advanced translation flows (e.g., tone-aware translation) * Refreshed LID and transliteration notebooks * Integrated with latest SDK structure #### Cleanup and Refactoring Removed deprecated `analytics` and `parse` APIs from SDK for better maintainability. --- # March 2025 ### Text #### Translation & Transliteration Schema Update Translation and Transliteration APIs will now automatically detect the input language. * Responses now include `source_language_code` * Improves handling of multilingual and code-switched inputs * Enables simplified workflows with less manual language tagging ### SarvamParse #### SarvamParse "Small" Mode Introduced A lightweight variant of SarvamParse is now available. * Lower cost * Faster response time * Ideal for real-time or cost-sensitive applications ### Language Identification #### LID API Released The new Language Identification (LID) endpoint detects both **language** and **script** from raw input text. * Supports multiple Indian and international languages * Detects script types like Latin, Devanagari, Kannada, and more * Full API Reference: [LID API Docs](https://docs.sarvam.ai/api/endpoints/language-identification) * Cookbook: [Language Identification Notebook](https://docs.sarvam.ai/api/api-guides-tutorials/text-processing/language-detection) --- # February 2025 ### APIs #### Sarvam Parse API Released A new API that transforms PDF documents into structured data. * Accepts PDF input and returns base64-encoded XML * Useful for information extraction, content indexing, and document analysis * API Reference: [Sarvam Parse Docs](https://docs.sarvam.ai/api/endpoints/sarvam-parse) * Cookbook: [Parse PDF Notebook](https://github.com/sarvamai/sarvam-ai-cookbook/tree/main/getting-started/doc-intelligence) #### Doc Translate API Released Translate full PDF documents and receive structured translated output. * Returns translated content as base64-encoded XML * Ideal for cross-lingual access to documents in enterprise, government, or education * API Reference: [Doc Translate Docs](https://docs.sarvam.ai/api/endpoints/doc-translate) * Cookbook: [Doc Translate Notebook](https://github.com/sarvamai/sarvam-ai-cookbook/tree/main/getting-started/doc-intelligence) ### Speech to Text (STT) #### Max Duration Limit Update To improve responsiveness and lower latency for all users, the maximum duration per STT request has been updated. * New limit: 30 seconds per request (previously 8 minutes) * Applies to both `saaras` and `saarika` models * For longer audio requirements, contact the team to explore tailored solutions --- # January 2025 ### Developer Experience #### Meta-Prompt Introduced Launched a universal **meta-prompt** to help guide any AI chat model in using Sarvam’s APIs effectively. * Offers structured API context for accurate prompt engineering * Compatible with models like Gemini, GPT, Claude, etc. * API Reference: [Meta-Prompt Docs](https://docs.sarvam.ai/api-reference/metaprompt) * Example: [Meta-Prompt in Action](https://shorturl.at/tpE4r) #### Sarvam AI Cookbook Launched Released the official **Sarvam AI Cookbook**, an open-source repository with practical code examples and notebooks. * Covers use cases for STT, TTS, Translation, Parse, and more * Includes best practices, integration tips, and tutorials * Repository: [Sarvam AI Cookbook on GitHub](https://github.com/sarvamai/sarvam-ai-cookbook) ---