Sarvam Prism

Sarvam's 105B decision model: typed questions about a state, a probability for every option in one pass, 44 ms median on a B200.

In short. Sarvam Prism is a decision model built on Sarvam's 105B language model. Instead of writing text, it takes a situation and questions with fixed answer options, and returns a probability for each option. It is built for automating routine decisions, it is strong on Indian languages, and it needs a very large GPU.

Summary

Sarvam Prism is a decision model that Sarvam AI released on 3 November 2026, built on its 105B mixture-of-experts language model with about 10.3B active parameters. A request gives a state, which can be any text or JSON, and one or more typed questions: choice, yes or no, or a score on ordered levels. The model returns a probability for every option in a single forward pass and generates no text, so the answer is always one of your options. A small decision head reads the model's hidden states, each option is encoded so that it cannot see the other options, and choice options are scored in a fixed order so the answer does not depend on the order you send them in. The context limit is 131,072 tokens per question.

Sarvam compared it with 24 other decision models on three indices that it ran itself: the 37 public benchmarks of the Jev Decision Index, its own recreation of the Artificial Analysis Decision Index and an internal Indian-language benchmark. Prism averages 71.8, with a median response of 44 ms on one B200 GPU, which Sarvam says is the highest score of any model that answers in under 100 ms. On the table in its post, mammoth-1 averages 75.2 and GLiDE 1.2 averages 73.2 but they take 284 ms and 116 ms, and Perplexity Decider v1.2 averages 68.0 at 117 ms. Prism scores 82.5 on the Indian-language index, against 75.0 for Perplexity Decider v1.2. Sarvam also reports that its confidence is calibrated, meaning its mistakes stay within the error it expects at 90% confidence or more. These are Sarvam's own results, on a benchmark mix it chose, and we have not reproduced them.

The weights use Apache 2.0, and Sarvam also offers a hosted API. They are stored in a custom FP8 layout, 106 GB in all, so they cannot be loaded with AutoModelForCausalLM or served by general-purpose inference servers: you use Sarvam's prism.py inference code on one NVIDIA GPU with FP8 support and about 120 GB of free memory, such as a B200 or H200. Sarvam says enterprises can train the calibrated base on their own decision data. In the 30 days to 11 November 2026, Hugging Face counted about 300 downloads.

Good for

  • Routing, triage and approval steps where an agent must choose among fixed options, with the probabilities used for thresholds
  • Decisions about text in Indian languages, where Sarvam reports its largest lead on its own index
  • Teams with a B200 or H200 that want a calibrated base to train on their own decision data

Watch out for

  • It needs one FP8-capable NVIDIA GPU with about 120 GB of free memory and only runs through Sarvam's own inference code
  • It cannot explain itself or take actions, and its probabilities should be tested against your own labeled examples
  • Every score is Sarvam's own, on a mix of indices it chose, including one it built itself, and two models in its table average higher but are slower
  • The model is new and has had few downloads so far, so there is little independent testing

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Sarvam Prism (custom FP8) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Sarvam Prism (custom FP8) FP8 105B 10.3B active 241 GB 121 GB 120 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 120 GB of GPU memory.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

Frequently asked questions

What is Sarvam Prism?
Sarvam Prism is a decision model built on Sarvam's 105B language model. Instead of writing text, it takes a situation and questions with fixed answer options, and returns a probability for each option. It is built for automating routine decisions, it is strong on Indian languages, and it needs a very large GPU.
Can I use Sarvam Prism commercially?
Yes. Sarvam Prism is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Sarvam Prism open source?
Sarvam Prism is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Sarvam Prism on my own computer?
Only on a machine with a lot of memory. The smallest variant, Sarvam Prism (custom FP8), needs about 120 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was Sarvam Prism released?
Sarvam Prism was released by Sarvam AI on November 3, 2026.
How long a context window does Sarvam Prism have?
Sarvam Prism supports a context window of up to 128K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Sarvam 30B and 105B Sarvam AI 32.1B – 106B 10.3B active 128K Apache 2.0 (Permissive) 22 GB Gaming GPU
pplx-decider Perplexity 4.54B – 26.1B 256K Apache 2.0 (Permissive) 3.13 GB Laptop or phone
AutoTrust Decider AutoTrust AI Lab 8.95B – 36B 3B active 256K Apache 2.0 (Permissive) 6.18 GB Laptop or phone
Clef Cloudflare 9.41B – 27.4B 64K Apache 2.0 (Permissive) 6.49 GB Laptop or phone