Sarvam Prism is an Apache 2.0 decision model built on Sarvam 105B

Sarvam has released Sarvam Prism, a decision model built on its Sarvam 105B mixture-of-experts model. Like the other decision models we have covered, it does not write text: you send a context and one or more questions with fixed answer options, and it returns a probability for every option in one forward pass. Sarvam’s launch post on 3 November says the weights are on Hugging Face under Apache 2.0, and we confirmed the repository’s LICENSE file is the Apache License 2.0.

How it is built

Sarvam says Prism is a LoRA fine-tune of Sarvam 105B with a small scoring head on top. The state and question are read together, each option sees that block and itself but never the other options, and only the final answer marker sees all of them. Training mixes cross-entropy with 5% label smoothing and a Brier loss, which Sarvam credits for the calibration. Options are scored in a fixed internal order (sorted by key), so shuffling them does not change the answer.

The scores, as Sarvam reports them

Sarvam ran Prism and 24 other decision models through three indices on its own hardware:

• Average: 71.8, from 67.0 on the Jev Decision Index (37 public benchmarks), 65.9 on Sarvam’s recreation of the Artificial Analysis Decision Index (with its private knowledge test replaced by Sarvam’s own hard questions) and 82.5 on an internal Indian-language benchmark.
• Speed: a median of 44 ms per request on one NVIDIA B200, which Sarvam says is the highest score of any model it tested under 100 ms. The model card puts the 95th percentile at 420 ms.
• Not the top score: in Sarvam’s table, Base Labs’ mammoth-1 averages 75.2 at 284 ms and GLiDE 1.2 averages 73.2 at 116 ms.
• Calibration: on English answers given with at least 90% confidence, Sarvam says Prism expects 3.3% errors and makes 2.2%; in Indian languages it expects 5.4% and makes 4.4%.

Sarvam’s API docs add a limit to that last point: in Indian languages, answers clear a 0.9 confidence threshold about 38.4% of the time, against 68.5% in English, so more of them need human review. Prism reads text only, in English, 10 Indian languages and Hinglish.

Running it

The model card asks for one NVIDIA GPU with FP8 support and about 120 GB of free memory, such as a B200 or H200, and about 106 GB of disk. The weights use a custom FP8 layout loaded by Sarvam’s prism.py, so they “cannot be loaded with AutoModelForCausalLM or served by general-purpose inference servers.” Locally the context limit is 131,072 tokens per question. The hosted Decisions API caps each question at 32,768 tokens and, per Sarvam’s docs, charges ₹4 per million input tokens with output free.

Why this matters

Prism joins open decision models such as Base Labs’ MIT-licensed mammoth-1 and elephant-1 and Cloudflare’s Apache 2.0 Clef, and its case is speed, calibration and Indian languages rather than the top average score. Unlike the RAIL-M terms on Nace AI’s Drex Max, Apache 2.0 leaves you free to fine-tune it on your own decisions, as Sarvam suggests, but self-hosting means a single large FP8 GPU and Sarvam’s custom code. If your routing or triage runs in Indian languages, test it at your real confidence threshold, since Sarvam’s own docs say fewer answers clear it there.