Sign in
Pending benchmarksTracking since Nov 3, 2026

Is Sarvam Prism state of the art?Pending benchmarks.Announced Nov 3, 2026; no independent results yet.

Sarvam AI's decision model built on Sarvam 105B: given a context and questions with fixed answer options, it returns a probability for every option in a single forward pass instead of generating text. English, 10 Indian languages and Hinglish. Open weights under Apache 2.0, plus Sarvam's API.

Announced facts

The details so far

Lab
Sarvam AI
Announced
Nov 3, 2026
API
Live · sarvam-prism · Sarvam API model sarvam-prism at POST /v2/systemone: ₹4 per 1M input tokens, output free
Open weights
Published · Hugging Face sarvamai/prism, not gated: a custom FP8 layout loaded with Sarvam's prism.py (about 106 GB); needs one FP8 GPU with about 120 GB free memory, tested on B200
Parameters
104,991,889,280 stored parameters (Sarvam 105B mixture-of-experts backbone with 128 routed experts per layer, plus a decision head)
Type
Decision model: one forward pass with a scoring head returns a probability per answer option; it does not generate text
Base model
LoRA fine-tune of Sarvam 105B
Question types
Choice, yes/no and score (ordered levels)
Training
Cross-entropy with 5% label smoothing plus a Brier loss
Context
131,072 tokens per question with the open weights; the hosted API accepts 32,768
Languages
English, 10 Indian languages (Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi) and Hinglish
Inputs
Text or JSON; no images, and no streaming in the API
License
Apache 2.0

Tags marked unsure are not stated on the lab's own pages.

Next steps

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

  • Publicly announced with a source
  • Usable by the public (API or weights)
  • Listed on an independent board we track
  • Enough results for a rank in a category
  • Holds across two snapshots (crowns only)
Self-reported

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: 71.8 average across three decision indices (Jev Decision Index 67.0, Sarvam's recreation of the Artificial Analysis Decision Index 65.9, an internal Indian-language benchmark 82.5), the highest average of the 25 decision models Sarvam tested that answer in under 100 ms; mammoth-1 (75.2) and GLiDE 1.2 (73.2) score higher but are slower.

Lab claim: 44 ms median per request (95th percentile 420 ms) on one NVIDIA B200.

Lab claim: at 90% or higher confidence, 2.2% of English answers and 4.4% of Indian-language answers are wrong, against 3.3% and 5.4% expected from the stated confidence.

Lab claim: 78.5% on judging agent runs (700 runs) and 89.0% on an internal security benchmark (5,517 questions); 0 of 2,000 answers change when the options are shuffled.

Receipts

Sources

LABSarvam AI announcement on Xx.com/SarvamAI/status/2117621415520448838 · Nov 3, 2026LABSarvam Prism (Sarvam blog)sarvam.ai/blogs/sarvam-prism · Nov 3, 2026LABSarvam docs: Sarvam Prismdocs.sarvam.ai/api/token-factory/models/sarvam-prismLABSarvam API reference: SystemOne decisionsdocs.sarvam.ai/api-reference/decisions/systemoneREPOsarvamai/prism on Hugging Facehuggingface.co/sarvamai/prism

Tracking continues until an independent board lists it.