Sarvam Prism Explained: India's 44 ms Open Decision Model

Indian AI company Sarvam AI has released Sarvam Prism, an open-weights "decision model" that does something different from ChatGPT or Gemini: it never writes a reply. You give it some context, a question and a list of options, and it returns a probability for each option in one pass. Sarvam says the median request takes 44 milliseconds on a single NVIDIA B200 GPU. Every benchmark below is Sarvam's own, so treat the numbers as a vendor claim until others test them.

Quick summary

• Released: November 3, 2026, by Sarvam AI (Bengaluru).
• Base: a LoRA fine-tune of Sarvam 105B, a mixture-of-experts model.
• Speed: 44 ms median per request on one B200 GPU.
• Score: 71.8 on average across three indices, the best of any model under 100 ms in Sarvam's 25-model test.
• Licence: open weights under Apache 2.0 on Hugging Face, plus a hosted API.
• Not for gaming PCs: it needs a GPU with at least 120 GB of memory.

What is Sarvam Prism, and what is a "decision model"?

Sarvam calls Prism a "System One" model, borrowing psychologist Daniel Kahneman's term for fast, intuitive judgement. Instead of generating text token by token, Prism reads a state (the situation), a question and the candidate options, then a small scoring head outputs a probability for every option. Think of it as a very fast multiple-choice engine rather than a chatbot.

That fits the jobs Sarvam names: routing customer-support tickets, document analysis, checking whether an AI agent's run succeeded, catching prompt-injection attempts, intent routing and picking which model should handle a request. Sarvam says it can take up to 131,072 tokens per question, so a long agent trace can be judged in one go.

How good is Prism, according to Sarvam?

Sarvam tested Prism against 24 other decision models on three indices: the public Jev Decision Index (37 benchmarks), its own recreation of an Artificial Analysis decision index, and an internal Indian-language benchmark covering ten Indian languages plus Hinglish. Sarvam timed the models it could run on the same B200 hardware.

So Prism is not the top scorer overall. Two models score higher but take roughly 2.6 to 6.5 times as long to answer. On Indian languages, Prism scored 82.5, ahead of Perplexity Decider (75.0), Jev 1.14 (70.8) and Microsoft-Decision-1 (70.0) in Sarvam's table.

Can you trust Prism's confidence?

Sarvam puts most weight on calibration: when a model says it is 90% sure, it should be wrong no more than about 10% of the time. For answers given with at least 90% confidence, Sarvam reports Prism expected 3.3% errors in English and actually made 2.2%. In Indian languages it expected 5.4% and made 4.4%. In the same test the other models made 2.7 to 4.8 times more mistakes than their confidence predicted on Indian languages.

Prism also sorts options into a fixed order before scoring. Sarvam says that left its answer unchanged on all 2,000 test questions when the options were shuffled, a known weakness of many AI judges.

Is Prism faster and cheaper than a normal chatbot as a judge?

For judging recorded AI-agent runs, Sarvam reports 78.5% accuracy for Prism against 70.7% for Claude Haiku 5.5 on the same 700 runs. Prism scored a run in about half a second on one GPU, while Sarvam measured Haiku 5.5 at a median 2.5 seconds per run through its API, at about $0.64 per 1,000 runs. On an internal cybersecurity test (phishing, jailbreaks, attack classification) Prism scored 89.0%. Sarvam says it will release that test publicly in the coming weeks.

Speed came from engineering: Prism started at 664 ms per request on two B200 GPUs in 16-bit precision, and Sarvam cut it to 44 ms by moving to FP8 on a single GPU, a grouped expert kernel, CUDA Graphs and merging the adapter into the weights.

How can you use Sarvam Prism in India?

Developers have two routes. The weights are on Hugging Face under the Apache 2.0 licence, which allows commercial use. Running it yourself needs an FP8-capable GPU with about 120 GB of free memory, such as a B200 or H200, so it is a data-centre model, not something for a gaming laptop or an RTX card. The easier route is the Sarvam API on Sarvam Indus. One secondary report lists the price at ₹4 per million input tokens; check Sarvam's pricing page before you build on that figure.

For everyday users, nothing changes today: Prism is plumbing for companies. It matters because it shows Indian labs competing on a specific, fast, auditable slice of AI rather than copying a chatbot. If you follow open models, compare it with Mistral Large 4's open weights and the Tencent Hy4 launch, and see where Google stands in our Gemini 4 Carbon explainer.

Verdict: why Prism is worth watching

Prism is a narrow tool with a clear pitch: near-top accuracy at a fraction of the latency, with honest confidence scores and the weights in your hands. The caveat is that all the evidence so far is Sarvam's own testing, including a benchmark it built itself. Independent results will decide whether the 44 ms claim and the calibration numbers hold up.

Related articles

Kimi K3.1 Pro Weights Released: 2.8T Parameters ExplainedMistral Large 4 Open Weights: Benchmarks and What It NeedsTencent Hy4 Launched: Vision, Computer Use, Price ExplainedAnthropic FlashProve Explained: Proving Which AI Answered