Ordalin
Submit a tool

Reviewed tool profile

Sarvam Prism

Sarvam AI's open-weight decision model that returns calibrated probabilities for yes/no, choice and score questions.

Sources checked

Visit website ↗
Pricing
The open weights are free to download under Apache 2.0 for self-hosting on your own GPU. The hosted Sarvam API charges ₹4 per 1M input tokens and nothing for output, the same on every plan, because Prism returns probabilities rather than text. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
Available as
Hosted API: the Decisions API (POST /v2/systemone), called with a Sarvam API key, with Python and JavaScript SDK examples; synchronous JSON with no streaming. · Self-hosted: weights on Hugging Face, run with the bundled prism.py code on one NVIDIA GPU with FP8 support and about 120 GB of free memory (such as a B200 or H200), plus about 106 GB of disk. Official source 1 ↗ Official source 1 ↗ · Official source 2 ↗
Primary group
Automation & Agents

Also consider

See all 6 similar tools ↓

What it does

Sarvam Prism is a System-1 decision model from Sarvam AI, released on 3 November 2026 and built on Sarvam 105B. Instead of writing text, it takes an input (the "state") plus typed questions (yes/no, pick-one or ordered score) and returns a calibrated probability for every option in one call. The weights are open under Apache 2.0 on Hugging Face, and the model is also served through Sarvam's Decisions API. Sarvam reports a median 44 ms per request on one NVIDIA B200 GPU.

Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗

Features

  • Typed questions, no text to parse. Each request carries a state and any number of yes/no, choice or score questions; the answer is always one of your options with a probability for each, so there is no output format to enforce. Official source 1 ↗
  • Calibrated confidence. On answers given with at least 90% confidence, Sarvam reports 2.2% actual errors against 3.3% expected in English and 4.4% against 5.4% in Indian languages, so confident answers can be automated and uncertain ones sent to a person. Official source 1 ↗
  • Same answer in any order. Choice options are sorted into a fixed internal order before scoring, so shuffling the options does not change the result. Official source 1 ↗
  • 44 ms median response. Sarvam measured a median 44 ms per request (95th percentile 420 ms) on one NVIDIA B200 and reports it as the highest-scoring model under 100 ms of the 25 decision models it tested. Official source 1 ↗ · Official source 2 ↗
  • English, Indian languages and Hinglish. Handles English, 10 Indian languages and Hinglish; Sarvam reports 82.5 on its internal Indian-language benchmark. Official source 1 ↗
  • Text or JSON input. The state can be plain text or any JSON object or array; images are not supported. Official source 1 ↗
  • Open weights you can fine-tune. Released with open weights under Apache 2.0 so enterprises can train it on their own data and run it on their own infrastructure. Official source 1 ↗ · Official source 2 ↗

Pricing & free limits

Open weights (self-hosted)

FreeApache 2.0 licence

Download from Hugging Face and run on your own NVIDIA GPU with FP8 support and about 120 GB of memory; you pay for your own hardware. Official source 1 ↗ · Official source 2 ↗

Sarvam API (sarvam-prism)

₹4 per 1M input tokensUsage-based, prepaid credits

Output tokens cost ₹0 and the price is the same on every plan. The state is counted once per request, so several questions about one input bill little more than the input itself. Official source 1 ↗ · Official source 2 ↗

Billing details

Only input tokens are billed; output is ₹0 because Prism returns probabilities, not text. Official source 1 ↗

The state is counted once per request however many questions you ask, so batching questions about the same input is cheaper; 3 short questions about a 1,000-token state cost about ₹0.004. Official source 1 ↗

Free access & limits

Free access and its limits were not established by the checked official sources.

Best for & limitations

Routing and triage, for example deciding which team should handle a support ticket and whether it is urgent, in place of prompting a chat model and parsing its reply. Official source 1 ↗

Judging recorded agent runs: whether the task succeeded, whether the run needs human review and which agent caused a failure. Official source 1 ↗ · Official source 2 ↗

Security guardrails such as prompt-injection and jailbreak checks, and moderation. Official source 1 ↗

  • Text only, no streaming. Prism reads text only and refuses images, and every request returns one JSON response without streaming. Official source 1 ↗
  • Heavy hardware to self-host. Self-hosting needs one FP8-capable NVIDIA GPU with about 120 GB of free memory, and the custom FP8 weights cannot be loaded with AutoModelForCausalLM or general-purpose inference servers. Official source 1 ↗
  • API request limits. On the API each question is limited to 32,768 tokens and 2–255 options, and a request that takes over 60 seconds fails with no answers; the self-hosted code allows 131,072 tokens per question. Official source 1 ↗ · Official source 2 ↗
  • Fewer confident answers in Indian languages. Only about 38.4% of Indian-language answers clear a 0.9 confidence threshold, against 68.5% in English, so more of them go to a person. Official source 1 ↗
  • Vendor-run benchmarks. The scores and speed comparisons were run by Sarvam itself, and probabilities can vary slightly between identical requests. Official source 1 ↗ · Official source 2 ↗
Sarvam Prism product screenshot
Website previewFixed desktop viewport · 1440 × 900

Similar tools

Published tools that share its categories.

OpenBotDesktop app that runs a team of persistent AI agents on your own computer with the Codex, Claude, Gemini or Grok plan you have.
Automation & AgentsFree planChecked Nov 4
Visit
MoA Mac on-screen assistant that does everyday tasks you ask for, and waits for your yes before each change.
Automation & AgentsPaidChecked Nov 1
Visit
JIMMYAn AI agent on your Windows PC that you control from your own WhatsApp chat to find and send files, run tasks and draft emails.
Automation & AgentsPaidChecked Oct 31
Visit