SSyncDev

Sarvam Prism

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Sarvam Prism is a 105B decision model that returns calibrated probabilities for yes/no, choice and score questions in English and 10 Indian languages.

Key specifications

Specification score
21.8
Context window
131.07K
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: INR 4 per 1M input tokens, INR 0 output, on every Sarvam plan; no USD price published. Indicative; confirm in your Sarvam account.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input

Produces

  • Text output

Capabilities

Open weights

What is Sarvam Prism?

Sarvam Prism never writes a sentence. You hand it a state (a support ticket, a transcript, a JSON record) plus a set of typed questions, and it returns a probability for every option of every question from a single forward pass. Classification, routing, scoring and guardrail checks are the jobs. Generation is not.

Released on 3 November 2026, it is a LoRA-style fine-tune of Sarvam 105B with a small decision head on top, and it covers English, ten Indian languages and Hinglish.

Request shape in brief

Calls go to POST /v2/systemone on the Sarvam API with model: "sarvam-prism". The body has two required parts:

  • state: text, an object or an array (objects are serialised as compact JSON);
  • questions: a map of your own IDs to question objects.

Three question types exist. noul is yes/no and returns the probability of yes. choice picks one of 2 to 255 options and returns the top option, a confidence and the full distribution. score takes 2 to 10 ordered levels, lowest first, and returns the expected level as a float. Generation fields such as temperature, max_tokens and stream are rejected, and images are not supported.

Option order does not matter for choice: when options were shuffled, 0 of 2,000 answers changed. For score, order defines the scale, so it matters a great deal.

Context: two different limits

The open weights are described as handling 131,072 tokens per question. The hosted API caps each question, meaning state plus question text plus options, at 32,768 tokens, and it never truncates the state; an oversized question fails the whole request with a 422. If you call the API, design for 32K. The larger figure applies when you serve the weights yourself.

Price and latency

Item Value
Input INR 4 per 1M tokens, every plan
Output INR 0 (no tokens generated)
Median latency about 44 ms per request on one B200
Request timeout 60 seconds

No USD price is published. Because the state is counted once per request, batching several questions about the same record is cheaper than separate calls. Figures are indicative; check your own Sarvam dashboard.

Self-hosting is not drop-in

Weights are Apache 2.0, but they ship in a custom FP8 layout loaded through Sarvam's prism.py. General-purpose inference servers cannot run them. Plan on about 120GB on a single FP8-capable GPU such as a B200 or H200.

Calibration is the selling point

Sarvam reports that for answers at 90% confidence or higher, observed error was 2.2% against 3.3% expected across 37 English benchmarks, and 4.4% against 5.4% in Indian languages. At a 0.9 threshold roughly 68.5% of English answers, but only about 38.4% of Indian-language answers, qualify for automatic action. That gap is the number to check on your own data. Label a few hundred real cases, pick a threshold, and measure how many decisions you could automate safely.

Other options live under structured decision models and the Sarvam provider page.

How to evaluate Sarvam Prism

Compare the model on the work you need to complete.

Tasks and inputs

Sarvam Prism is listed for Text classification, Zero-shot classification, Structured Decisions. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 131.07K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Calibrated probabilities for each option of each question (yes/no, choice or score). It generates no text, so output tokens are always zero.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Google

Gemini Decision 1

Gemini Decision 1 powers Google's Decisions API, returning calibrated yes/no, choice and rubric answers instead of text, at $0.10 per 1M input tokens.

262.14K context

Perplexity

pplx-decider-flash-v1.3

pplx-decider-flash-v1.3 is Perplexity's small, half-price decision model, distilled from the 27B decider, for high-volume classification and span extraction.

— context

Perplexity

pplx-decider-v1.3

pplx-decider-v1.3 is Perplexity's 27B-class decision model: it returns probabilities for choice, yes/no, score and new span-locate questions, not text.

32.77K context

Butler

Cicero

Cicero is Butler Compute's open-weight MoE decision model: up to 64 typed questions per state, a calibrated probability for every option, no text.

131.07K context

Morph

Reflex

Morph Reflex is a family of fast classifier models that label agent turns (jailbreak, looping, frustration and more) with per-class scores.

65.54K context

levanto

Sage 2

Levanto Sage 2 is listed as a new version of Levanto's hosted Sage decision model; Levanto's own docs do not yet name a version 2.

— context