India’s Sarvam updates its 105B open-weight model for enterprise agents — at a fraction of rivals’ price

India’s Sarvam AI released an updated, openly downloadable version of its flagship 105-billion-parameter AI model on 22 November, tuned for enterprise agent work — and priced, the company says, at roughly a fifth of a comparable Google model’s rate. Moneycontrol reported the release on 22 November, with a same-day write-up on the trade outlet Unite.AI.

The release, Sarvam-105B-1711, is an update to the company’s 105B model, not a new foundation model. Its base is a mixture-of-experts model — a design in which only a small, specialised slice of the network is used for any given request — with 105 billion parameters in total but 10.3 billion active per request. The Apache-2.0 open-source weights include a compressed 4-bit build that the company says fits on a single 96 GB RTX PRO 6000 Blackwell GPU.

The post-training — the extra tuning stage that shapes a base model into a usable assistant — targets tool calling, instruction following, agentic tasks and Indian languages.

The benchmark gains are the company’s own figures and have not been independently verified: instruction following on IFBench rose from 32.5 to 82.5, and customer-service tool use on the Tau2 benchmark from 46.2 to 85.5. On Tau2’s pass^4 measure — the share of tasks completed correctly on all four attempts — Sarvam reports 69.5 per cent, against 63.9 per cent for Google’s Gemini 3.5 Flash-Lite. The blended price — a single average rate across input and output tokens — is $0.15 per million tokens, down from $0.44 for its predecessor and against $0.85 for Gemini 3.5 Flash-Lite. Unite.AI notes the Indic language scores come from Sarvam’s internal benchmark, judged by another AI model, Gemini 3.1 Pro.

Sarvam also reports infrastructure gains: continued pretraining on 330 billion tokens extended the context window — the text the model can hold in view at once — to 160,000 tokens.

The release lands mid-controversy. Sarvam’s Token Factory, an inference service launched in mid-November that runs other companies’ open models — including China’s DeepSeek — on Indian servers, has stirred a debate in India about whether renting out foreign models advances or hollows out the country’s AI mission (Indian tech-policy outlet Medianama covered the debate on 19 November; Economic Times reported the launch). Moneycontrol framed the new release as coming “amid Token Factory criticism” — and the company has pushed back. Co-founder Pratyush Kumar told Moneycontrol that offering inference does not conflict with Sarvam’s core mission, and Sarvam has clarified that Token Factory runs entirely on privately procured GPUs, with requests never forwarded to foreign APIs. The criticism is a policy debate rather than an allegation of wrongdoing.

Verification note: Moneycontrol’s article is login-walled; we worked from its headline and standfirst plus Unite.AI’s detailed report (an AI-assisted but editor-reviewed trade outlet). All performance and price figures are company-reported and unverified.

Why it matters: The single-GPU footprint and low price target firms that want AI agents inside their own networks rather than behind a foreign API — the exact deployment India’s sovereignty debate is about. Whether Sarvam’s self-reported gains hold up in production, not just in its own benchmarks, will decide the model’s uptake.

Suggested internal links:
– “how AI models are evaluated” → post 16400 (OpenAI’s perplexity-based development limits)
– “independent verification of AI claims” → post 16383 (Riemann proof claim and its verification timeline)

The post India’s Sarvam updates its 105B open-weight model for enterprise agents — at a fraction of rivals’ price appeared first on DailyAI.