Sarvam Prism: an open decision model in 44 ms

What happened?

On November 3, India's Sarvam AI released Prism, an open model trained specifically for decisions (Sarvam). Sarvam calls it a "System One" model: instead of reasoning at length, it returns a probability for every answer option in a single forward pass. It is built for classification, routing and fast checks inside agents.

How it works

Prism is a LoRA trained on Sarvam's 105B mixture-of-experts model (105B parameters, 10B active). A question can be up to 131,072 tokens. The output is not text but calibrated probabilities for each option, so the score can be used as a threshold.

Key takeaways

• An average score of 71.8 with 44 ms median latency on a single B200. Sarvam says that is the best result among 25 tested models under 100 ms.
• 82.5 on Indian languages.
• Calibration: at over 90% confidence, actual error was 2.2% versus 3.3% expected in English, and 4.4% versus 5.4% in Indian languages.
• Shuffling the option order changed the decision in 0 of 2,000 cases.
• On agent trace evaluation it scored 78.5% versus 70.7% for Claude Haiku 5.5, which took 2.5 seconds.
• Latency fell from 664 ms to 44 ms, roughly a 15x speedup.

Availability

The weights are available under Apache 2.0 in FP8 on Hugging Face. Running it needs a GPU with at least 120 GB of memory, and it is also offered through the Sarvam API.

Why it matters

Agent systems make many small decisions: which tool to call, whether a task is done, whether a step is safe. Using a slow, expensive frontier model for these is wasteful. A fast, calibrated and permissively licensed decision model can serve as a cheap guard or router, though its 120 GB memory requirement means it still needs serious hardware.