Ga direct naar de inhoud
zondag 22 november 2026Gratis gids
NL
dailyai
Onderzoek

India’s Sarvam updates its AI model at a fraction of rivals’ price

India’s Sarvam AI has released an updated version of its flagship AI model — one that anyone can download and run themselves — tuned for business tasks such as agents that work unsupervised. The company prices it at roughly a fifth of what a comparable Google model costs, Moneycontrol reported.

· Leestijd: 2 minuten
DeelXin

India’s Sarvam AI has released an updated version of its flagship AI model — one that anyone can download and run themselves — tuned for business tasks such as agents that work unsupervised. The company prices it at roughly a fifth of what a comparable Google model costs, Moneycontrol reported.

The release, called Sarvam-105B-1711, updates the company’s existing model rather than starting fresh. It uses a design where only the small slice of the network needed for each request is switched on, which keeps it cheap to run: about 10.3 billion of its 105 billion total parameters — the learned settings that make a model work — are active at any one time.

Anyone can download and reuse the model free under a standard open-source licence. The company says a squeezed-down copy fits on a single high-end desktop computer.

The extra training for this release focused on tool use, following instructions, acting on the user’s behalf, and Indian languages.

The gains are the company’s own figures and zijn niet onafhankelijk geverifieerd. Its score for following instructions went from 32.5 in February to 82.5, and for a customer-service agent task it went from 46.2 to 85.5.

On a tougher test — the share of tasks completed correctly in four separate tries — Sarvam reports 69.5 per cent, against 63.9 per cent for Google’s Gemini 3.5 Flash-Lite. The price is $0.15 per million tokens — the units a model reads and writes, like characters of work — down from $0.44 for its predecessor and against $0.85 for Google’s model.

One caveat: the scores for Indian languages come from Sarvam’s own test, marked by another AI model, Unite.AI notes. Sarvam also says extra training extended the model’s “context window” — how much text it can hold in view at once — to 160,000 tokens.

The release lands mid-controversy. Sarvam’s Token Factory service, opened earlier this month, runs other companies’ open models — including China’s DeepSeek — on Indian servers. That has stirred a debate in India about whether renting out foreign models advances or hollows out the country’s AI mission (Medianama covered the debate on 19 November).

The company has pushed back. Co-founder Pratyush Kumar told Moneycontrol that running other firms’ models does not conflict with Sarvam’s core mission, and Sarvam has clarified that Token Factory runs on privately procured GPUs, with requests never forwarded to foreign APIs. The criticism is a policy debate rather than an allegation of wrongdoing.

Waarom het belangrijk is: The small-computer footprint and low price are aimed at firms that want AI agents running inside their own systems rather than through a foreign company’s service — exactly the deployment India’s sovereignty debate is about. Whether Sarvam’s self-reported gains hold up in real use, not just in its own tests, will decide how widely the model is taken up.