Sarvam Prism
by Sarvam AI
Not yet tested by us. Every score and speed on this page is vendor-reported by the model's maker or by Sarvam, as linked. Facts checked 3 Nov 2026.
What is Sarvam Prism? Sarvam Prism is a decision model that Sarvam AI released on 3 Nov 2026. It is a LoRA fine-tune of Sarvam 105B that reads an input plus typed questions and returns a probability for every answer option in one pass, without generating text. Sarvam publishes the weights under Apache 2.0 (per Sarvam's blog, 3 Nov 2026).
Editor note
An open-weight decision model from Sarvam AI, released 3 Nov 2026 and built on Sarvam 105B. It returns a probability for each option of yes/no, pick-one and rubric-score questions instead of writing text, which suits judging agent runs and routing. Weights are on Hugging Face under Apache 2.0; it is also a paid API.
- What it tests
- Not a benchmark: a decision model you can use as an eval judge or router in your own pipeline.
- Original source
- docs.sarvam.ai/api/token-factory/models/sarvam-prism
- Last updated
- Released 2026-11-03
- License
- Apache-2.0
Prism, Jev, Clef and Perplexity Decider for judging and routing (all figures vendor-reported)
| Model (maker) | What it decides / tests | Weights and licence | Price (vendor list) | Sarvam's 3-index average (vendor-reported by Sarvam) | Agent-run judging accuracy, 700 runs (vendor-reported by Sarvam) | Speed (vendor-reported) |
|---|---|---|---|---|---|---|
| Sarvam Prism (Sarvam AI) | Yes/no, pick-one and rubric-score questions on text or JSON; Sarvam pitches routing, triage, moderation, judging agent runs and security checks. English, 10 Indian languages and Hinglish. | Open weights, Apache 2.0 (Sarvam blog) | ₹4 per 1M input tokens, output free (Sarvam docs) | 71.8 | 78.5% | 44 ms median on one B200 GPU (Sarvam) |
| Jev 1.14 (TypeSafe AI) | Choice, Score and Noul (yes/no) questions on a text or JSON state; API only. | Closed, API only | $0.042 per 1M input tokens, output free (TypeSafe launch post) | 64.5 | 76.4% | 196 ms as reported publicly; 431 ms when Sarvam called it from its cluster |
| Clef (Cloudflare) | Noul, choice and score questions on text and images, Jev-compatible API, up to 64 questions per request. | Open weights, Apache 2.0 (Cloudflare) | $0.24 per 1M input tokens on Workers AI (Cloudflare) | 62.6 | 76.4% | 209 ms median (Cloudflare); 222 ms (Sarvam) |
| Perplexity Decider v1.2 (Perplexity) | Choice, noul and score questions on text, JSON and images; Perplexity suggests it for rubric grading, routing and moderation. | Weights on Hugging Face (licence not checked by us) | $0.02 per 1M input tokens, output free (Perplexity docs) | 68.0 | Not in Sarvam's agent-run test | 117 ms (Sarvam) |
When to use Sarvam Prism
- As an eval judge: you want a pass/fail or rubric score on agent runs with a probability you can threshold, and you need a lot of them cheaply. Sarvam reports judging a full agent trace in about half a second on one GPU (vendor-reported).
- As a router: you need a fast pick-one decision (which team, which model, which tool) with a confidence value, especially for Indian-language or Hinglish traffic, where Sarvam's own index puts Prism ahead (vendor-reported).
- You need to self-host: the weights are open, unlike Jev's API-only model.
Limitations
- Every number here comes from Sarvam's own launch post, where Sarvam built two of the three indices itself (a recreation of the Artificial Analysis Decision Index and an internal Indian-language benchmark). We have not reproduced any of it.
- Self-hosting needs one GPU with at least 120 GB of memory in FP8, per Sarvam. The 44 ms figure was measured on an NVIDIA B200.
- Text and JSON only; Sarvam's docs say images are not supported. Clef and Perplexity Decider accept images.
- Context limit: Sarvam's blog says up to 131,072 tokens per question, while its API docs list 32,768 tokens per question. Check which applies to your route.
- It does not write explanations. If your judge must say why a run failed, keep a text-generating model for that step.
How to run it
- Get a Sarvam API key and install the SDK (sarvamai 0.1.43 or later for Python, per Sarvam's docs).
- Send a state plus noul, choice or score questions to the Decisions API with model sarvam-prism.
- Read the probability per option and act only above a threshold you set on your own labelled sample.
- Or download sarvamai/prism from Hugging Face and run it yourself.
pip install -U "sarvamai>=0.1.43"
# Sketch from Sarvam's docs:
from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key=<key>)
resp = client.decisions.create(model="sarvam-prism", state=<text>, questions={...})Notable results
- Vendor-reported: 71.8 average across three indices at 44 ms; the highest average under 100 ms of the 25 decision models Sarvam tested. Two models (mammoth-1, GLiDE 1.2) scored higher but responded more slowly. [Sarvam AI blog: Introducing Sarvam Prism (3 Nov 2026)]
- Vendor-reported: at 90%+ confidence on English questions, Prism expected 3.3% errors and made 2.2%. Sarvam reports Jev 1.14 made 2.7 times as many errors as its confidence implied. [Sarvam AI blog: Introducing Sarvam Prism (3 Nov 2026)]
- Vendor-reported: with its fixed-order step, Prism changed 0 of 2,000 answers when options were shuffled; Sarvam measured Jev 1.14 changing 12.3% and Perplexity Decider v1.2 10.3%. [Sarvam AI blog: Introducing Sarvam Prism (3 Nov 2026)]
- Vendor-reported: 89.0% on Sarvam's internal 5,517-question security benchmark, against 84.5% for Clef and 78.6% for Jev 1.14. The benchmark is not public yet. [Sarvam AI blog: Introducing Sarvam Prism (3 Nov 2026)]
Sources
- Sarvam AI blog: Introducing Sarvam Prism (3 Nov 2026)
- Sarvam docs: Sarvam Prism model page
- Hugging Face: sarvamai/prism
- TypeSafe AI: Introducing System One models and Jev (15 Sep 2026)
- Cloudflare changelog: Introducing Clef (1 Oct 2026)
- Cloudflare blog: Clef-omni and Clef pricing (9 Oct 2026)
- Perplexity docs: Decisions API
- Hugging Face: perplexity-ai/pplx-decider-v1.2
Related on Evals Review
- Jev (TypeSafe AI): the decision model that started the category
- Jev vs Perplexity Decider: those two compared on price, inputs and on-device use
- LLM-as-judge tools: every judge and rubric tool we track
FAQ
What is Sarvam Prism?
A decision model from Sarvam AI, released 3 Nov 2026 and built on Sarvam 105B. It answers yes/no, pick-one and rubric-score questions with a probability per option rather than generating text.
Sarvam Prism vs Jev: which is better?
On Sarvam's own benchmarks Prism averages 71.8 against 64.5 for Jev 1.14 and responds faster, but these are vendor-reported and partly on indices Sarvam built. Jev is API-only; Prism's weights are open. Test both on your own labelled data.
What are Sarvam Prism's benchmark results?
Vendor-reported by Sarvam: 71.8 average over three indices (67.0 Jev Decision Index, 65.9 on its Artificial Analysis recreation, 82.5 Indian languages) and 78.5% on judging 700 recorded agent runs.
Is Sarvam Prism open source?
Sarvam says the weights are on Hugging Face under the Apache 2.0 licence.
Can I use Sarvam Prism as an LLM judge?
Yes for closed questions: did the task succeed, does this run need review, which rubric level fits. It cannot write a rationale, so pair it with a text model if you need one.
Related rubric libraries & frameworks evals
- Inspect Evals — Many benchmarks under one runner: coding, agents, safety, knowledge.
- OpenAI Evals — Building and running your own evals; community registry.
- LLMJury rubric library — LLM-as-judge rubrics for support, RAG, summaries, code and tone.
- Prompt Assay judge templates — Starter LLM-as-judge prompts.
Discussion
No one has discussed Sarvam Prism yet. Have you run it? Share what it caught, what it missed or how you set it up.
Related reading
- Jev and Perplexity Decider head to head: Jev vs Perplexity Decider
- The model Prism is most often measured against: Jev (TypeSafe AI)
- All judge and rubric tools: LLM-as-judge tools