Sarvam Prism
AvailableSarvam's "System One" decision model, released Nov 3 2026 with Apache 2.0 open weights on Hugging Face (sarvamai/prism) and on the Sarvam API. A LoRA fine-tune of the Sarvam-105B mixture-of-experts base (128 routed experts) with a small decision head: given a state, a question and a fixed set of options, it returns a calibrated probability for each option in a single forward pass and does not generate text. Context limit 131,072 tokens per question. Weights ship in a custom FP8 layout for the bundled prism.py inference code (one GPU with about 120 GB, such as a B200 or H200); Sarvam reports a 44 ms median response and a 71.8 average across three decision indices, the best under 100 ms among 25 decision models it tested (vendor-reported).
Specifications
- License
- Open source · Apache-2.0
- Weights
- Downloadable
- Architecture
- Mixture-of-Experts
- Parameters
- 105B
- Context window
- 131K tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- —
- Modalities
- Text
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.