Sarvam AI models
Bengaluru company that trains speech, translation, document vision and language models for Indian languages and offers them through its API, open weights and agent products.
Bengaluru, IndiaFounded 2023sarvam.ai
- Releases
- 21
- Open weights
- 9
- Latest
- Nov 03, 2026
November 2026
1 releaseSarvam Prism
Sarvam Prism is a decision model fine-tuned from Sarvam 105B that scores a fixed set of answer options in a single forward pass and returns a probability for each one instead of generating text.
- Modalities
- text → score
- Context
- 128K context
- Parameters
- MoE: 105B total, 10B active
- License
- Apache-2.0
- At launch
- Open weights, API
- Sarvam reported an average of 71.8 across three decision indices at a 44 ms median response on one NVIDIA B200, the top score under 100 ms among the 25 decision models it tested
- Built as a LoRA fine-tune of Sarvam 105B with a small decision head, answering choice, yes/no and ordered-score questions
- Weights were published on Hugging Face under Apache 2.0 in a custom FP8 format that runs on one GPU with about 120 GB of memory
- At launch it was also offered as a hosted model on the Sarvam API; the model card lists a limit of 131,072 tokens per question
October 2026
1 releaseBulbul V4 Flash
Bulbul V4 Flash is Sarvam's real-time text-to-speech model, distilled from larger Bulbul V4 research models to keep their expressive delivery in Indian languages at lower latency and cost.
- Modalities
- text → audio
- At launch
- API
- Launched with 83 voices across 11 languages and 224 voice and style combinations
- Sarvam reported under 250 ms time to first audio in production
- Priced at launch at ₹20 per 10,000 characters, against ₹30 for Bulbul V3
- Offered at launch through the Sarvam API, Playground and Content Studio; the same post previewed a larger Bulbul V4 Pro model
September 2026
1 releaseSarvam Vision 2.1
Sarvam Vision 2.1 is an updated version of Sarvam's document intelligence vision-language model that adds key-value extraction from forms, stronger multi-page table parsing and handwriting recognition for Indian languages.
- Modalities
- text, image → text
- At launch
- API
- Sarvam reported an overall olmOCR-Bench score of 87.3 on the official set, the highest in its comparison table
- Post-trained with supervised fine-tuning and reinforcement learning with verifiable rewards on new synthetic and real form and handwriting data
- Kept the earlier design of a vision-language model wrapped by a semantic layout parser and a reading order network
- At launch it was offered through the Document AI Digitise and Extract APIs and the Akshar document agents; Sarvam had previewed these features as Vision 2.0 at its Epoch event
August 2026
2 releasesSaaras V4
Saaras V4 is Sarvam's speech recognition model for 22 Indian languages and English, pairing an audio encoder with a 3B hybrid state-space decoder so one model can return five transcript formats.
- Modalities
- audio → text
- At launch
- API
- Output modes cover verbatim and normalized transcription, code-mixed text, transliteration and translation
- Added single-pass multi-speaker recognition that combines diarization and transcription
- Sarvam reported the lowest word error rate of the systems it tested on seven English test sets, and a 5.22% language identification error across 22 Indian languages
- Announced at Sarvam Epoch; the changelog added saaras:v4 to the Speech-to-Text API on August 18, 2026, Voice Agents made it generally available on September 2, and a technical post followed on September 25
Sarvam 105B Conversations
Sarvam 105B Conversations is a variant of Sarvam 105B post-trained for real-time dialogue, voice agents and multi-turn chat in Indian languages and English.
- Modalities
- text → text
- Context
- 32K context
- At launch
- API
- On a production benchmark built from Indian customer conversations, Sarvam reported a voice-call score of 82.1, against 74.7 for GPT 5.4 Mini and 78.9 for Gemini 3.5 Flash
- Served under the model ID sarvam-105b-conversations with a 32K-token context window, against 128K for the base model
- Introduced at Sarvam Epoch together with a reasoning-tuned Sarvam 105B update for agent work, both offered through the Sarvam API
- Sarvam's August 18, 2026 changelog priced it the same as sarvam-105b
March 2026
2 releasesSarvam 105B
Sarvam 105B is Sarvam's flagship mixture-of-experts reasoning model, trained from scratch in India for reasoning, coding and agentic work with a focus on Indian languages.
- Modalities
- text → text
- Context
- 128K context
- Parameters
- MoE: 105B total, 10.3B active
- License
- Apache-2.0
- At launch
- Open weights, API, Consumer app
- About 10.3B active parameters per token, 128 experts and Multi-head Latent Attention; pre-trained on 12T tokens
- Sarvam reported 98.6 on Math500, 71.7 on LiveCodeBench v6 and 90.6 on MMLU
- Weights were released under Apache 2.0 on Hugging Face and AI Kosh, with access through the Sarvam API and the Indus assistant
- Trained on compute provided under the IndiaAI Mission; it had powered the Indus beta that opened on February 20, 2026
Sarvam 30B
Sarvam 30B is a mixture-of-experts reasoning model trained from scratch by Sarvam and tuned for fast, low-cost deployment such as multilingual voice calls with tool use.
- Modalities
- text → text
- Context
- 64K context
- Parameters
- MoE: 30B total, 2.4B active
- License
- Apache-2.0
- At launch
- Open weights, API
- Retired
- 2.4B active parameters per token, 128 experts and grouped-query attention; pre-trained on 16T tokens
- Sarvam reported 88.3 on AIME 25, or 96.7 with tools, and 66.5 on GPQA Diamond
- Weights were released under Apache 2.0 on Hugging Face and AI Kosh, and the model was offered through the Sarvam API
- It already powered Samvaad, Sarvam's conversational agent platform, at release
Sarvam fully deprecated sarvam-30b on its Chat Completions API on August 18, 2026 and pointed users to Sarvam-105B; the open weights stayed on Hugging Face.
Source: Open-sourcing Sarvam 30B & 105B models | Sarvam AI
February 2026
6 releasesSarvam Edge
Sarvam Edge is a set of compact on-device models for speech recognition, speech synthesis and translation in Indian languages, designed to run on phones and laptops without a network connection.
- Modalities
- text, audio → text, audio
- Parameters
- 74M ASR, 24M TTS, about 150M translation
- Speech recognition model with 74M parameters, about 294 MB in FP16, covering 10 Indian languages with automatic language detection
- Text-to-speech model with 24M parameters, about 60 MB, covering 10 Indian languages with 8 multilingual speakers
- Translation model of about 150M parameters covering 11 languages and 110 direct language pairs
- At launch Sarvam said it was building Edge with global device manufacturers and listed no public download or API
Saaras V3
Saaras V3 is Sarvam's speech recognition model for 22 Indian languages and English, built on a new architecture with native streaming for low-latency transcription.
- Modalities
- audio → text
- At launch
- API
- Sarvam reported a 19.31% word error rate on the 10 most-used languages of the IndicVoices benchmark, against about 22% for Saaras V2.5
- Trained on more than 1 million hours of curated multilingual audio through pre-training, supervised fine-tuning and reinforcement learning
- Decodes incrementally for real-time transcription and also supports batch processing
- At launch it was offered through the Saaras V3 API and a no-code experience in the Sarvam dashboard
Bulbul V3
Bulbul V3 is Sarvam's third-generation text-to-speech model, built on an LLM that infers emphasis, pauses and pacing from context to produce natural speech in Indian languages.
- Modalities
- text → audio
- At launch
- API
- Launched with more than 35 professional voices across 11 languages, plus voice cloning and a low-latency streaming mode
- In a blind A/B listening study run by Josh Talks across 11 languages, Sarvam reported it was the most preferred model for 8 kHz telephony audio
- Sarvam said language support would later expand to 22 Indian languages
- Sarvam gave all users unlimited API access to the model through February 28, 2026
Sarvam Vision
Sarvam Vision is a 3B-parameter state-space vision-language model for document intelligence, OCR and general image understanding in English and 22 Indian languages.
- Modalities
- text, image → text
- Parameters
- 3B
- At launch
- API
- Built by continued pre-training of Sarvam's 3B base model, followed by supervised fine-tuning and reinforcement learning with verifiable rewards
- Handles image captioning, scene text, charts and complex tables, with a layout parser and a reading order network around the model
- Sarvam reported leading results on olmOCR-Bench (English subset) and OmniDocBench v1.5, and introduced a 22-language Sarvam Indic OCR Bench with 20,267 samples
- The Document Intelligence API and dashboard experience were free to use for all of February 2026
Sarvam Audio
Sarvam Audio is a speech recognition model built as an audio extension of Sarvam 3B, a language model trained from scratch on English and 22 Indian languages, that uses conversation context to improve its transcripts.
- Modalities
- audio, text → text
- Lets applications choose one of five output styles at inference time, from literal transcription to code-mixed, romanized or English-translated text
- Sarvam reported state-of-the-art diarized speech recognition on recordings of 1 to 60 minutes with up to 8 speakers
- Can turn speech directly into a tool call with arguments for voice agents
- At launch Sarvam said it would be available soon on the Sarvam Dashboard
Sarvam Dub
Sarvam Dub is a dubbing model that moves speech between Indian languages while keeping the original speaker's voice and fitting the new audio to the video's timing.
- Modalities
- audio → audio
- At launch
- Enterprise
- Uses zero-shot voice cloning and builds duration control into speech generation rather than stretching audio afterward
- Sarvam reported state-of-the-art speaker similarity on its benchmark of 64 speakers across 10 Indian languages and English
- Used to dub the monthly Mann Ki Baat address into 11 languages and to dub the Union Budget 2026 speech live into Kannada and Hindi
- Sarvam reported a 6.6x latency reduction over a base PyTorch implementation for live use
June 2025
1 releaseSarvam-Translate
Sarvam-Translate is an open-weights translation model, fine-tuned from Gemma 3 4B with AI4Bharat, that translates whole documents between English and the 22 official Indian languages.
- Modalities
- text → text
- Context
- 8K context
- Parameters
- 4B
- License
- GPL-3.0
- At launch
- Open weights, API
- Handles long-form and structured input such as Markdown, HTML, code and scientific notation while keeping the formatting
- Sarvam reported that language experts rated it well above Gemma3-27B-IT, Llama 4 Scout and Llama-3.1-405B-FP8 in human evaluation
- Weights were released on Hugging Face under GPL-3.0, and the model was offered through Sarvam's API
- The model card lists a maximum context of 8K tokens
May 2025
1 releaseSarvam-M
Sarvam-M is a 24B hybrid reasoning model that Sarvam post-trained from Mistral Small with supervised fine-tuning and reinforcement learning to improve Indian language, math and programming performance.
- Modalities
- text → text
- Context
- 8K context
- Parameters
- 24B
- License
- Apache-2.0
- At launch
- Open weights, API
- Retired
- One model with switchable think and non-think modes
- Sarvam reported average gains over the base model of 20% on Indian language benchmarks, 21.6% on math and 17.6% on programming
- Weights were released on Hugging Face under Apache 2.0, with access through Sarvam's playground and API
- Sarvam called its post the first in a series of technical write-ups to accompany planned regular model releases
Sarvam deprecated sarvam-m on its Chat Completions API on June 5, 2026 and directed users to Sarvam-30B and Sarvam-105B; the weights stayed on Hugging Face.
Source: Sarvam-M multilingual language model | Sarvam AI
October 2024
1 releaseSarvam-1
Sarvam-1 is a 2B-parameter base language model trained from scratch for 10 Indian languages and English, with a tokenizer that needs far fewer tokens per word for Indic scripts.
- Modalities
- text → text
- Context
- 8K context
- Parameters
- 2B
- License
- Sarvam AI Research License
- At launch
- Open weights
- Trained on about 4 trillion tokens, including the 2-trillion-token Sarvam-2T Indic corpus, on 1,024 GPUs of Yotta's Shakti cluster over five days
- Tokenizer fertility of 1.4 to 2.1 tokens per word across the supported languages
- Sarvam reported it outperformed Gemma 2 2B and Llama 3.2 3B on MMLU, ARC-Challenge and IndicGenBench
- Released on Hugging Face as a text-completion model meant for fine-tuning, under a non-commercial research license
September 2024
1 releaseMayura
Mayura is Sarvam's API translation model for English and Indian languages, trained on formal and colloquial data so its output reads like everyday speech, including code-mixed text.
- Modalities
- text → text
- At launch
- API
- Covered 10 Indian languages at launch and handled respectful second and third person forms
- Offered a gender toggle for first-person translations and kept formatting, URLs, codes and emojis intact
- Could break long multi-clause English sentences, such as legal text, into shorter and more readable Indic sentences
- Sarvam's translation API was part of its August 13, 2024 platform launch; this post introduced the model under the Mayura name
August 2024
3 releasesBulbul v1
Bulbul v1 was Sarvam's first text-to-speech model, a single compact multilingual model built for code-mixed Indian speech.
- Modalities
- text → audio
- At launch
- API
- Launched with six voices that keep the same identity across languages, in more than 10 languages
- Trained to pronounce domain terms, local names, numbers and abbreviations common in Indian usage
- Supported pitch and pace control for every voice
- Sarvam's speech synthesis API was part of its August 13, 2024 platform launch; this post named and detailed the model
Sarvam 2B
Sarvam 2B was Sarvam's first language model trained from scratch, a small base model for 10 Indian languages and English released as an early pre-trained checkpoint.
- Modalities
- text → text
- Parameters
- 2B
- License
- Sarvam AI Research License
- At launch
- Open weights
- The Hugging Face card described the v0.5 checkpoint as pre-trained on 2 trillion tokens with no post-training
- Sarvam's tokenizer table showed an average of 2.08 tokens per word across supported languages, against 9.34 for Llama 3.1
- Trained with NVIDIA NeMo on Yotta's Shakti cloud using HGX H100 systems
- Announced at Sarvam's August 13, 2024 platform launch; the fully trained version shipped as Sarvam-1 in October 2024
Shuka v1
Shuka v1 is an open audio language model that understands spoken questions in Indian languages by joining Sarvam's Saaras v1 audio encoder to Meta's Llama 3 8B Instruct.
- Modalities
- audio, text → text
- License
- Llama 3 Community License
- At launch
- Open weights
- The encoder and decoder are linked by a projector of about 60M parameters, the only part that was trained
- Trained on less than 100 hours of audio, with the projector tuned on English and Hindi data
- Sarvam tested it on 11 languages; beyond English and Hindi, it answered spoken questions through zero-shot transfer from the multilingual encoder
- Announced as Shuka 1.0 at Sarvam's August 13, 2024 platform launch and released on Hugging Face
December 2023
1 releaseOpenHathi-7B-Hi-v0.1-Base
OpenHathi-7B-Hi-v0.1-Base was Sarvam's first open model, a 7B Llama 2-based base model extended for Hindi, English and Hinglish.
- Modalities
- text → text
- Parameters
- 7B
- License
- Llama 2 Community License
- At launch
- Open weights
- First release in Sarvam's OpenHathi series, built with academic partner AI4Bharat
- Added 16K Hindi tokens to the Llama 2 tokenizer for a 48K vocabulary to cut token counts on Hindi text
- Released as a base model meant to be fine-tuned rather than used directly
- Published on Hugging Face under the Llama 2 license
Cite and reuse this data
These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.
Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".
Citation
The AI Engineering Editors. "Sarvam AI models: release history and timeline." The AI Engineering, updated Nov 03, 2026. https://theaiengineering.com/releases/sarvam-ai
HTML link
<a href="https://theaiengineering.com/releases/sarvam-ai">Sarvam AI models: release history and timeline</a> (The AI Engineering)- Full dataset
- 1,250 releases from 73 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering
Other labs
Google106
Alibaba Qwen84
OpenAI79Mistral AI59
Meta44
- Z.ai43
Microsoft42Tencent42
- ByteDance Seed40
- DeepSeek38
- Liquid AI34
Cohere31- IBM30
xAI30- NVIDIA29
MiniMax28
Stability AI28- Anthropic26
Ai223StepFun22
Amazon21- Baidu21
- ElevenLabs21
Black Forest Labs18
- Kuaishou16
- Meituan16
Runway16- Moonshot AI15
Midjourney14
BharatGen13Salesforce13
- Xiaomi13
Nous Research11- Perplexity10
AI4Bharat9
Krutrim9
AI21 Labs8- Apple8
Ideogram8
01.AI7
Hugging Face7
Lightricks7Luma AI7
Reka7- Suno7
Gnani AI6- Databricks5
Leonardo.Ai5- Pika5
Recraft5- Mannas4
Soket AI4
Fractal3
Inflection AI3MBZUAI IFM3
- Snowflake3
Thinking Machines Lab3- Infosys2
NPCI2
Tech Mahindra2
TypeSafe AI2
ZenteiQ2ARTPARK (IISc)1
Avataar1- Cognition1
CoRover.ai1
Genmo1Josh Talks1
Reflection AI1
Ricursive Intelligence1Shunya Labs1
- Zoho1