Reported News
Sarvam Prism: an open-weights decision model on Sarvam 105B that returns probabilities, not text
Sarvam Prism, released by Sarvam AI on November 3, 2026, is a decision model built on Sarvam 105B: you send an input and questions with fixed options, and it returns a probability for every option. Its model card says "It does not generate text, so the answer is always one of your options." Weights are Apache 2.0.
- Changes reported
- 5
- listed below, each from a source we read
- We tested
- Not tested yet
- reported from sources; we did not run it ourselves
- Gaps we flag
- 4
- things we could not verify
- Sources
- 3
- 3 primary, linked below
What changed
| # | What the source says | Status |
|---|---|---|
| 01 | Sarvam AI's release post (Nov 3, 2026) introduces Prism as a "System One" decision model trained on the Sarvam 105B-10B base, which the post ties to mixture-of-experts trade-offs. | Reported |
| 02 | It answers yes/no, pick-one and ordered-scale questions about a text or JSON input with a probability for every option, in one call; images are not supported (hosted API docs). | Reported |
| 03 | Weights are on Hugging Face (sarvamai/prism) under Apache 2.0. Running them needs one NVIDIA GPU with FP8 support and about 120 GB of free memory, such as a B200 or H200, and about 106 GB of disk. | Reported |
| 04 | The hosted API (model ID sarvam-prism, POST /v2/systemone) costs ₹4 per 1M input tokens with output not billed, and allows 32,768 tokens per question and 2 to 255 options per choice question. | Reported |
| 05 | Sarvam reports a 71.8 average across three indices at a 44 ms median response on one B200, which it calls the highest score under 100 ms among the 25 decision models it tested; it scores 82.5 on Sarvam's internal Indian-language benchmark. | Reported |
reported by the sourcemeasured or run by usnot verified
Primary source: Sarvam AI blog: Introducing Sarvam Prism (Nov 3, 2026). We are the fast, independent index; the vendor’s own page is the source of record for this release.
Who it matters for
Teams that today prompt a chat model to "answer yes or no" or "pick one of these labels" and then parse the reply: ticket routing, triage, moderation, prompt-injection checks and judging agent runs. Prism returns a typed answer with a probability, so your code sets the threshold for acting on its own. Sarvam's agent-trace test (700 recorded runs) puts Prism at 78.5% accuracy against 70.7% for Claude Haiku 5.5, on Sarvam's own setup.
Sarvam credits TypeSafe's Jev with starting the decision-model category last month; Prism adds an Apache 2.0 option that enterprises can fine-tune on their own decision data. Self-hosting is a data-center job, though: one B200 or H200 class GPU. Teams without that hardware would use the hosted API, priced in rupees.
In the source’s words
“It does not generate text, so the answer is always one of your options.”
“We are releasing Prism with open weights so that enterprises can train this calibrated base on their own data and run it on their own infrastructure.”
What we have not verified
- We have not run Prism, locally or through the API.
- All benchmark, speed and calibration numbers are Sarvam's own; it ran the 25 models through its indices itself, one index is Sarvam's recreation of the Artificial Analysis Decision Index and another is internal.
- The release post says the model handles up to 131,072 tokens per question; the hosted API docs list 32,768. We report the API limit.
- Pricing is as listed in Sarvam's API docs on November 4, 2026.
This brief is Reported: written from the sources below, not from running the release ourselves.
Sources
- Sarvam AI blog: Introducing Sarvam Prism (Nov 3, 2026)primary source · fetched 2026-11-04
- Sarvam-Prism model card, Hugging Faceprimary source · fetched 2026-11-04
- Sarvam API docs: Sarvam Prismprimary source · fetched 2026-11-04
Related
Spotted an error? See our corrections policy or email the address on the contact page. Markdown version for agents: /news/sarvam-prism-open-weights-decision-model.md
Want these briefs in your Google Top Stories? Add Agentic Weekly as a preferred source on Google.
Newsletter
The Weekly Verdict. Free, every Monday.
A free email every Monday: the week's few big AI releases, then 5 short use cases for the new models, each with a prompt you can paste.
Free. Double opt-in. One click to unsubscribe.