Summary:
Sarvam’s new decision model is available through an API and open weights. Its benchmark results are promising, but deployment hardware, test methods and confidence thresholds matter.Sarvam AI, an Indic-language AI company, released Sarvam Prism on November 3. Rather than write a conversational answer, the model assigns probabilities to predefined choices. That makes it a candidate for the small classification decisions inside a larger application, not a replacement for a general-purpose chatbot.
Available through an API or open weights
The official model repository lists Apache 2.0 licensing and a hosted API option. Local deployment needs an FP8-capable NVIDIA GPU with about 120 GB of free memory and roughly 106 GB of disk space. Its custom weight layout requires the supplied inference code rather than a generic text-generation server.
According to the API documentation, the model ID is sarvam-prism. Requests contain text or JSON and questions with yes/no, choice or ordered-scale answers. Images are not supported. The hosted API accepts up to 32,768 tokens per question and returns synchronous JSON, without streaming. Access requires a Sarvam API account and key; open weights do not mean hosted inference is free.
How Prism compares with selected decision models
The following scores come from Sarvam’s own evaluation, not independent testing. Bold numbers identify the strongest result within each displayed row; the shaded Prism column identifies the subject, not a blanket winner.
| Benchmark (higher is better) | Sarvam Prism | Perplexity Decider v1.2 | Jev 1.14 | Clef | Microsoft-Decision-1 |
|---|---|---|---|---|---|
| Jev Decision Index | 67.0 | 67.7 | 60.7 | 61.7 | 57.8 |
| Sarvam’s AA recreation | 65.9 | 61.2 | 61.9 | 54.8 | 55.6 |
| Internal Indian-language test | 82.5 | 75.0 | 70.8 | 71.2 | 70.0 |
| Three-index average | 71.8 | 68.0 | 64.5 | 62.6 | 61.1 |
Methodology: the Jev index contains 37 public benchmarks. Sarvam recreated the Artificial Analysis index with a replacement for its private knowledge test; the Indian-language test is internal. These are selected peers, not a whole-market ranking. The recreation must not be presented as an official Artificial Analysis leaderboard result.
What the speed claim means
The model card reports a 44 ms median and a 420 ms 95th-percentile response time across 1,000 sequential requests on one NVIDIA B200. Those figures are not a hosted-API latency guarantee. Sarvam’s broader comparison mixes measured local timings with reported timings for API-only peers, so it is not a controlled end-to-end speed ranking.
A practical way to evaluate it
Start with a narrow, reversible task such as sorting incoming requests into predefined queues. Build a held-out set with the labels your team actually uses, including ambiguous examples. Compare both mistakes and the amount of work sent for human review at each confidence threshold.
A fixed answer format removes parsing failures, but it does not make the selected answer correct. High-confidence mistakes should still be measured separately. For costly or consequential actions, retain a review step rather than treating a benchmark result as permission to automate the decision.
Visuals: original data graphics by PromptsZone, not AI-generated imagery. Sarvam wordmark © Sarvam AI, used under its brand guidelines; no endorsement implied. Sources checked November 4, 2026.
