Everything we announced

Sarvam 105B

We released the first Sarvam 105B checkpoint in February. At Epoch, we released updated variants for conversational agents and work agents, the two use cases where we are seeing the most traction in India.

For conversational agents, standard public benchmarks are incomplete. They do not capture ambiguous messages, ASR noise, code-mixed speech, long policy documents, or instructions that must be followed many turns later.

We therefore evaluated the model on a production benchmark built from hundreds of millions of minutes of Indian conversations.

Sarvam 105B leads GPT 5.4 Mini and Gemini 3.5 Flash on all four measures, with the widest gaps in instruction following and linguistic quality, the two capabilities that decide whether a voice agent survives a real customer call. It does this at $0.80 per million blended tokens, against $4.50 for GPT 5.4 Mini and $9.00 for Gemini 3.5 Flash. Better on every measure, at a fraction of the price.

For work agents, we tuned a reasoning variant and evaluated it on HarnessBench: 106 sandboxed tasks across software engineering, data and finance, knowledge retrieval, office workflows, tool use, SRE, and long-running autonomy.

Against GLM 5.2, a 753B model, Sarvam 105B performs close enough that the most useful product decision is routing. In our own agent configuration, 36% of tasks route to Sarvam 105B and 64% to GLM 5.2, reducing serving cost by roughly 40%.

Both variants are available through the Sarvam API: one optimised for low-latency conversations, the other for complex tasks inside an agent harness.

Sarvam Vision 2.0

Sarvam Vision 1.0 launched in February and has since digitised more than 35 million pages.

Vision 2.0 extends the model with native key-value extraction from tables and forms, Indic handwriting recognition, stronger complex-table parsing, and improved general OCR.

We also released Sarvam Vision Edge to bring these capabilities into customer environments through models customised for specific workflows. In a land-record digitisation workflow for the Government of Odisha, a custom model achieved 50% higher accuracy than the best general-purpose model tested on the same task.

Saaras V4

Saaras V4 extends Sarvam's speech-recognition stack with five output modes from one model: verbatim transcription, normalised transcription, code-mixed output, transliteration, and translation.

It also adds multi-speaker ASR in a single pass, combining diarisation and transcription rather than running separate sequential systems. This is useful for call recordings, meetings, and interviews, where speaker attribution matters as much as the transcript.

Saaras V4 supports 22 Indian languages. And the best English ASR is now an Indian one: across seven standard suites, from LibriSpeech and GigaSpeech to Svarah for Indian-accented speech, it records the lowest average word error rate of any system tested.

Bulbul V4 (Research preview)

Bulbul V4 is our most expressive text-to-speech model yet. Developers can shape how each line is spoken, guiding emotion, emphasis, pacing, and cues such as laughter as the audio is generated.

This brings more control to voice agents, creator tools, dubbing, education, accessibility, and media. It is especially valuable in enterprise settings, where names, amounts, product terms, and tone must be delivered clearly and accurately.

Custom Model Training

We're introducing Custom Model Training for organisations that want models specialised on their own data, policies, and workflows. We manage training on infrastructure in India and deliver the trained weights to the organisation, bringing our experience in managing GPU fleets efficiently and running optimised model training pipelines.

We work with organisations to train smaller, task-specific models that can reduce latency and inference costs compared with using a larger general-purpose model for every request.

Throughout the engagement, the organisation's training data stays isolated and is never used to train Sarvam's own models. Once training is complete, the resulting weights are delivered to and owned by the organisation, which retains full control over how the model is deployed and modified.