Sarvam open-sources Prism, a 44ms decision model that skips text entirely
heyshrutimishra · x · 2026-11-04
Indian AI company Sarvam released Prism, an open-weights model built for fixed-option decisions rather than text generation: give it a question with candidate options (approve/reject, ticket routing, prompt-injection detection) and it returns a probability per option in a single pass — the answer is always one of your choices, nothing to parse.
Key numbers:
- 71.8 average across three decision benchmarks, which Sarvam claims is the highest among 25 sub-100ms decision models tested; 82.5 on an Indian-language benchmark
- 44ms median latency on one NVIDIA B200 — 15x faster than the 664ms first version, with no accuracy loss
- Consistent answers across 2,000 questions with shuffled options; at 90%+ stated confidence, measured error rates were 2.2% (English) and 4.4% (Indian languages)
Works in English, 10 Indian languages, and Hinglish. Weights are on Hugging Face under Apache 2.0; the API costs ₹4 per million input tokens. Benchmarks are Sarvam's own, but a fast, cheap, open model for boring yes/no decisions targets exactly the classification work most enterprises actually need.
Related event: India's Sarvam Open-Sources Prism, a 44ms System One Decision Model(6 posts)→
More from Venture
- Developer hits 3 clients in one week demanding GEO over SEO — daluoseo · 2026-11-04
- ByteDance Seedance overseas revenue up ~5x in six months; Tencent and Bilibili among top clients — lxfater · 2026-11-04
- Meta's earnings call: Muse hit 10M+ setups in ~6 weeks, ~$200 average purchase — SuB8u · 2026-11-04
- Shanghai field trip: China investors say the AI trade is over, question 2027 capex — SumitGup · 2026-11-04
- Enterprise AI spend hits $177B, up 5.4x; Anthropic and OpenAI take 86% of API spend — deedydas · 2026-11-04
- Chart: CoreWeave burned $2.80 per $1 of revenue at IPO vs Anthropic's $0.14 — NathanpmYoung · 2026-11-04