# Sarvam AI — Introducing Saaras V4

- Company: Sarvam AI (sarvam.ai)
- Announced: 2026-09-25T12:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.sarvam.ai/blogs/introducing-saaras-v4/
- Record: https://forck.live/items/18382-introducing-saaras-v4
- Subject: Sarvam / Bulbul / Saaras
- Models affected: Saaras V4

Sarvam AI released Saaras V4, a speech recognition model with a 3B hybrid state-space language model decoder trained from scratch. It achieves state-of-the-art performance on all 22 Indian languages and the lowest word error rate on seven English benchmarks. The model supports five transcript formats (transcribe, translate, verbatim, translit, codemix) and has a language identification error rate of 2.9% across the top ten Indian languages.

## Evidence

Verbatim from https://www.sarvam.ai/blogs/introducing-saaras-v4/:

> Saaras V4 uses an audio encoder, and an LLM decoder. The LLM decoder is a 3B hybrid state space model completely trained from scratch in-house

---

Record: https://forck.live/items/18382-introducing-saaras-v4
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md