{"version":1,"type":"story","url":"https://digestai.news/story/sarvam-ai-releases-open-weight-prism-model-for-enterprise-decisions","json":"https://digestai.news/story/sarvam-ai-releases-open-weight-prism-model-for-enterprise-decisions.json","markdown":"https://digestai.news/story/sarvam-ai-releases-open-weight-prism-model-for-enterprise-decisions.md","slug":"sarvam-ai-releases-open-weight-prism-model-for-enterprise-decisions","headline":"Sarvam AI releases open-weight Prism model for enterprise decisions","summary":"Sarvam AI launched **Prism**, an open-weight decision model built on its 105B base, on November 3, 2026. Unlike generative models, Prism scores predefined options in one forward pass without token-by-token generation, targeting enterprise use cases like intent routing, document intelligence, and agent oversight. It achieves a **44-millisecond median response time** on a single NVIDIA B200 GPU (FP8 precision) and scores **71.8** on Sarvam’s internal benchmarks, outperforming competitors like mammoth-1 in latency while maintaining accuracy. The model supports **131,072 tokens per question** and is licensed under Apache 2.0, with weights hosted on Hugging Face and API access via Sarvam Indus. Sarvam optimized inference by fusing FP8 conversion, CUDA Graphs, and kernel grouping, claiming no accuracy loss from these changes.\n\nPrism’s design emphasizes calibration: it uses **cross-entropy with 5% label smoothing** and **Brier loss** to reduce overconfidence and penalize costly mistakes. Testing on **226,088 English questions** and **20,108 multilingual queries** showed error rates below predicted thresholds. For agent evaluation, it scored **78.5%** accuracy across 700 recorded runs, excelling in task success (75.0%) and failure attribution (69.6%). Sarvam also highlights **89.0%** accuracy on an internal cybersecurity benchmark of 5,517 questions, with plans to release the dataset publicly. The model’s open weights enable enterprises to deploy it on private infrastructure, addressing a gap in AI-driven decision automation.","keyPoints":["Prism scores predefined options in one pass, targeting enterprise decisions like intent routing and agent oversight","Achieves **44ms median latency** on a single NVIDIA B200 GPU (FP8) with **71.8** benchmark score","Open weights under Apache 2.0, hosted on Hugging Face; API access via Sarvam Indus"],"whyItMatters":"Prism fills a niche for enterprises needing fast, explainable decision-making without generative AI’s latency or cost. Its open weights and calibration-focused design could accelerate adoption in high-stakes workflows like cybersecurity and agent oversight, where confidence thresholds matter.","category":{"slug":"enterprise","name":"Enterprise & Industry","url":"https://digestai.news/category/enterprise"},"entities":{"companies":["Sarvam AI","NVIDIA","Hugging Face","Microsoft"],"models":["Prism","Sarvam 105B","mammoth-1","Jev 1.14","Microsoft-Decision-1","Clef"],"people":[]},"firstPublishedAt":"2026-11-04T11:11:02Z","updatedAt":"2026-11-04T11:11:02Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"Unite.AI","title":"Sarvam AI Releases Open-Weight Prism Decision Model","url":"https://unite.ai/sarvam-ai-releases-open-weight-prism-decision-model","publishedAt":"2026-11-04T11:11:02Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[],"thread":null,"cite":{"text":"Digest AI, \"Sarvam AI releases open-weight Prism model for enterprise decisions\", 4 November 2026, https://digestai.news/story/sarvam-ai-releases-open-weight-prism-model-for-enterprise-decisions","publisher":"Digest AI","title":"Sarvam AI releases open-weight Prism model for enterprise decisions","datePublished":"2026-11-04T11:11:02Z","url":"https://digestai.news/story/sarvam-ai-releases-open-weight-prism-model-for-enterprise-decisions"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}