Sarvam-Prism

Sarvam-Prism

Scores every answer option for questions about a text, with a probability for each, in one forward pass. For teams that need fixed-choice decisions, such as ticket routing, on a large NVIDIA GPU.

Made by sarvamai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Sarvam-Prism answers multiple-choice style questions about a piece of text. It sits on top of Sarvam-105B. For each question you list the allowed answers, and it returns a probability for every one of them in a single forward pass. It writes no free text, so the result is always one of your options.

Who it suits: It suits teams who turn text into fixed decisions, such as routing a support ticket or rating how upset a customer is. They need a large NVIDIA GPU or Sarvam's hosted API.

Industry | Who uses it | How they use it | Example
Customer support | Support teams | Route tickets and flag urgent ones | Pick billing, tech or delivery for a ticket
Operations | Teams triaging requests | Score how upset a customer is on an ordered scale | Rate a message from calm to furious
Software | Developers building decision steps | Replace free-text classification with fixed options | Add a yes or no check to a workflow

People who need generated text can skip it, since the model does not generate any.

Where to use it:

• Python on one NVIDIA GPU with FP8 support, such as a B200 or H200
• Sarvam's hosted API, which needs no GPU
• Windows, macOS or Linux, where you can download the files with Python

Use case | Who | What they get back
Ticket routing | Support teams | A probability for each team, with the most likely one named
Urgency check | Operations | A probability of yes for a yes or no question
Tone scoring | Analysts | An expected level and probabilities across ordered levels

How to set it up: You need these first:

• A single NVIDIA GPU that supports FP8, with roughly 120 GB of free memory. The card tested it on a B200
• Python 3.10 or newer, plus a C compiler, since Triton builds its kernels on first use
• Around 106 GB of free disk for the model files

1. Download the model and its code into a local folder.
2. Install the dependencies from the requirements file.
3. Run the example script on that folder.

pip install -U huggingface_hub
huggingface-cli download sarvamai/prism --local-dir sarvam-prism
cd sarvam-prism

pip install -r requirements.txt

python example.py .

How to use it: Load the model from the downloaded folder, then call decide with a state and a set of questions. On a B200 the first load runs about two minutes, since Triton builds its kernels and CUDA graphs get captured along the way.

from prism import Prism

model = Prism.from_pretrained(".")   # the downloaded folder; loads on cuda:0

answers = model.decide(
    state={"ticket": "My card was charged twice for the same order and I still haven't got a refund."},
    questions={
        "team": {"type": "choice", "instructions": "Which team should handle this ticket?",
                 "criteria": {"billing": "payments, charges and refunds",
                              "tech": "app or website problems",
                              "delivery": "shipping and order status"}},
        "urgent": {"type": "yes_no", "instructions": "Does this need a reply within the hour?"},
        "anger": {"type": "score", "instructions": "How upset is the customer?",
                  "criteria": ["calm", "annoyed", "angry", "furious"]},
    },
)

Tool | What it does
Prism.from_pretrained | Loads the model from the downloaded folder onto cuda:0
decide | Scores every question in a request together
choice | Gives the most likely key and a probability for each key
yes_no | Gives the probability of yes
score | Gives an expected level, the most likely level and probabilities

Example: You send a ticket saying a card was charged twice with no refund, and ask three questions. The card's sample output picks billing for the team with probability 0.986, gives 0.45 for yes on urgent, and scores anger at 1.86.

For newcomers and experts: Think of it as a multiple-choice grader for text. You write the question and the allowed answers, and it tells you how likely each answer is. A first small step is to run the example script and read the output.

A request can hold several questions. They are scored together, with up to 16 per batch by default, yet each one is judged separately. Choice options are always scored in key order, so sending them in a different order changes nothing. Each question can carry up to 131,072 tokens of context. The files use a custom FP8 format, so general-purpose inference servers and AutoModelForCausalLM cannot load them.

You send a context and questions
Each question lists its allowed answers
One forward pass scores every answer
You read a probability per answer

Good to know:

• Size: 105B parameters per Hugging Face, about 105 GB in F8_E4M3
• Languages: English plus Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati and Kannada are listed
• Speed: the card measured a median of 44 ms per request on one B200
• Loading: it captures CUDA graphs on load, and you can pass cuda_graphs=False to skip that

Title: Latency per request on one B200
Median: 44 ms
95th percentile: 420 ms

• License: apache-2.0
• Cost to run: no key or account for a local run, but it needs a GPU with about 120 GB of free memory. The hosted API is a separate Sarvam service
• Data it reaches: it downloads model files from Hugging Face, and the script reads only the text you pass as state
• Not for: writing text, or loading through AutoModelForCausalLM, because the files use a custom FP8 format

On Hugging Face, 13 November 2026

• Task: text classification
• Parameters: 105B
• Weights: about 105 GB in F8_E4M3; about 52.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 325
• Likes: 44
• License: apache-2.0
• Languages: en, hi, bn, ta, te, mr, gu, kn
• Fine-tunes: 1
• Tags: safetensors, sarvam_mla, decision-model, classification, calibration, text-classification

Links

• Model: https://huggingface.co/sarvamai/prism
• Files: https://huggingface.co/sarvamai/prism/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sarvamai/prism

Then follow the usage in its model card. Nothing in this kit runs by itself.