Sarvam Prism
Prism is a decision model built on Sarvam 105B. It returns probabilities for predefined choices rather than generating an answer string. Both hosted API access and Apache 2.0 weights are published.
The output contract is a decision
The Decisions API accepts yes/no, pick-one and ordered-scale questions. Each option receives a probability. The local release performs a single forward pass and documents approximately 106 GB of weight storage.Sarvam Prism Decisions API guideSarvam Prism official weights and requirements
Practical use
Practical guidance. Define labels that cover the real input distribution. Include a review route for uncertain cases, then measure calibration on your own labelled examples. A confidence value is useful only when it predicts errors in that workload.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | sarvam-prismsarvamai/prism |
| Context window | 32,768 tokens Per question, including state, question and options Sarvam Prism Decisions API guide |
| Input | Text or JSON; no image inputSarvam Prism Decisions API guide |
| Local hardware guidance | FP8-capable NVIDIA GPU with about 120 GB free memorySarvam Prism official weights and requirements |
Related models
- Sarvam 105B
Sarvam AI · Sarvam
A 105B mixture-of-experts reasoning model for Indian-language text, code and tool calls, with 10.3B active parameters.
Launch date not confirmedAPI · Downloadable