Sarvam AI has released Prism, an open-weights "System One" decision model built on its Sarvam 105B base. It does not write text. It reads a state, a question and a list of options, and returns a probability for each option in a single forward pass. Sarvam says the median request takes 44 ms on one NVIDIA B200 GPU and that Prism scores 71.8, the highest average among 25 decision models it tested that respond in under 100 ms.
The details come from Sarvam's launch post, published November 3, 2026. All benchmark numbers are Sarvam's own, run by Sarvam, so read them as a vendor claim until others replicate them. The weights, per Sarvam, are on Hugging Face under the Apache 2.0 licence.
TL;DR: what Prism is, in a table
| Question | Answer |
|---|---|
| What is it? | A decision model: picks among options and gives a probability for each |
| Base | Sarvam 105B (a 105B mixture-of-experts with about 10B active), fine-tuned with LoRA |
| Speed | 44 ms median on one B200, 15x faster than the team's first version (664 ms) |
| Headline score | 71.8 average over three indices, best under 100 ms among 25 models, by Sarvam's test |
| Higher-scoring models | mammoth-1 (75.2 in 284 ms) and GLiDE 1.2 (73.2 in 116 ms) |
| License | Apache 2.0 weights on Hugging Face, plus the Sarvam API |
| Hardware | FP8 on one GPU with at least 120 GB of memory |
| Who verified it? | Sarvam only so far |
What is a "System One" decision model?
Sarvam borrows the term from Kahneman's fast, intuitive thinking. Much of enterprise work is a chain of small judgements: route this ticket, flag this run, allow or block this prompt. A decision model is built for those. It answers with probabilities over a fixed set of options, so a workflow can act when confidence is high and hand off to a person when it is not.
Sarvam credits a startup, Typeface, with creating the category last month and naming its Jev model. In the same chart Sarvam compares Prism with Jev 1.14, Perplexity Decider, Microsoft-Decision-1, Clef and others. For context on Sarvam's wider lineup, see our guide to Sarvam's models and API and the Sarvam Epoch announcements.
The scores, and what to read into them
Sarvam averages three indices. The first is the Jev Decision Index, which it describes as 37 public benchmarks. The second is Sarvam's recreation of the Artificial Analysis Decision Index, where it swapped out the private knowledge test for curated hard questions with an "I don't know" option. The third is an internal Indian-language benchmark.
| Model | Average | Jev | AA recreation | Indian languages | Time (ms) |
|---|---|---|---|---|---|
| mammoth-1 | 75.2 | 75.3 | 70.2 | 80.1 | 284 |
| GLiDE 1.2 | 73.2 | 72.0 | 68.1 | 79.4 | 116 |
| Sarvam Prism | 71.8 | 67.0 | 65.9 | 82.5 | 44 |
| elephant-1 | 71.3 | 72.2 | 67.6 | 74.0 | 145 |
| Perplexity Decider v1.2 | 68.0 | 67.7 | 61.2 | 75.0 | 117 |
Source: Sarvam's launch post. Three honest observations:
- Prism is not the top scorer. mammoth-1 and GLiDE 1.2 score higher, but take two to six times as long.
- On the public Jev index alone, Prism's 67.0 trails Perplexity Decider (67.7) and several others. Its lead comes from the Indian-language set (82.5) and Sarvam's own recreation.
- Sarvam built two of the three indices, so the average tilts toward tests it controls.
Response time is the median of 1,000 requests sent one at a time, each model on its developers' own inference code. API-only models were plotted at their reported times, and Sarvam says measured from its cluster Jev 1.14 and Microsoft-Decision-1 took 431 ms and 533 ms.
Calibration: does the confidence mean anything?
The feature Sarvam stresses most is calibration. An enterprise acts only above a confidence threshold, which is safe only if the model's stated confidence matches its real error rate.
Sarvam checked expected versus actual mistakes on answers given with at least 90% confidence:
| Model | English: expected | English: actual | Indian languages: expected | Indian languages: actual |
|---|---|---|---|---|
| Sarvam Prism | 3.3% | 2.2% | 5.4% | 4.4% |
| Clef | 2.8% | 2.3% | 3.9% | 15.3% |
| Perplexity Decider | 1.4% | 2.0% | 3.0% | 8.1% |
| Microsoft-Decision-1 | 1.7% | 3.1% | 3.0% | 10.1% |
| Jev 1.14 | 1.3% | 3.5% | 3.4% | 16.1% |
Prism makes fewer errors than it predicts at every threshold from 50% to 95%, in both English (226,088 questions) and Indian languages (20,108 questions). Sarvam attributes this to two training choices: cross-entropy with 5% label smoothing, which discourages complete certainty, and a Brier loss, which makes confident mistakes costly.
Same answer in any order
Models that pick among options can change their answer when the options are shuffled. Sarvam tested 2,000 questions with four shuffles. Several rivals changed 9% to 12% of answers. Prism, reading options as sent, changed 4.0%. Sarvam then added a step that places the options in one fixed order before scoring and maps probabilities back to the original names. With that, 0 of 2,000 answers changed.
Where it could be used
Sarvam says it tested Prism on speech-to-speech harness decisions, trace evaluations, prompt-injection guardrails, intent routing for customer support, model routing and document intelligence. It shares two early benchmarks.
Judging agent runs
Using 700 recorded agent runs (1,770 questions) from AgentRewardBench, tau-bench, Who&When and TRAIL, with labels from human experts or the environment's own checks:
| Question | Prism | Clef | Jev 1.14 | Microsoft-Decision-1 | Claude Haiku 5.5 |
|---|---|---|---|---|---|
| Did the task succeed? | 75.0% | 71.6% | 71.8% | 71.3% | 66.0% |
| Needs human review? | 86.4% | 80.4% | 81.9% | 83.4% | 77.9% |
| Which agent caused failure? | 69.6% | 60.9% | 65.2% | 45.7% | 37.0% |
| All questions | 78.5% | 76.4% | 76.4% | 75.7% | 70.7% |
Sarvam says Prism scores a full run in about half a second on one GPU and handles up to 131,072 tokens per question. It says Claude Haiku 5.5 through its API takes a median 2.5 seconds per run and costs $0.64 per 1,000 runs. Prism's cost is not stated in comparable terms, so we cannot say which is cheaper at your volume.
Security guardrails
On an internal benchmark of 5,517 questions across nine tasks built from vulnerabilities published since July 2026 (phishing artifacts, jailbreak prompts, MITRE attack types), Prism scored 89.0%, versus 84.5% for Clef, 78.6% for Jev 1.14 and 76.1% for Microsoft-Decision-1. Sarvam says it will release the set publicly in the coming weeks. Until then, nobody outside Sarvam can check it. For other guardrail approaches, see our look at a local proxy that blocks risky coding-agent tool calls.
How it works under the hood
Prism is a LoRA fine-tune of Sarvam 105B. State, question and options go in together. A small scoring head returns a probability per option, with no token-by-token generation.
The key design is the attention mask. The state and question form one block read in both directions. Each option sees that block and itself, never the other options, and all options start from the same position. Only the answer token sees everything. Isolating options makes each judgement independent, and reading the block both ways adds about 3 points on the Jev index, per Sarvam.
Speed came from engineering. The first version took 664 ms per median request in 16-bit precision on two B200s. Sarvam converted weights to FP8 to fit one GPU, ran all 128 experts in one grouped kernel, captured the forward pass as a CUDA Graph, folded the FP8 conversion into the expert kernels and merged the adapter into the weights. It splits the mask computation so the dense block uses a standard fast kernel; that makes attention 2.6 times faster past 2,048 tokens. FP8 moved the Jev score by 0.01 points. Together the steps give the 15.1x speedup.
How to try it
- Weights: download from Hugging Face under Apache 2.0 (Sarvam's link is in its post). You need FP8 support and one GPU with at least 120 GB of memory.
- API: Prism is on the Sarvam API as Sarvam Indus, with documentation linked from the launch post.
- Input shape: each request is a state, a question and any number of options within the context window. The reply is a probability per option.
- Fine-tuning: Sarvam's pitch is that enterprises can train this calibrated base on their own decisions and run it on their own infrastructure.
Limits and open questions
- Vendor-run benchmarks. Sarvam ran every model, wrote one index and used an internal security set.
- Hardware cost. A 120 GB GPU is not small. Sarvam gives no price per decision.
- Narrow task. A decision model chooses among given options. It will not draft, plan or call tools.
- Calibration across domains. The calibration data covers public benchmarks and Sarvam's own set. Your data may differ, so re-check on a held-out sample before you automate.
- No outside replication yet.
Related Sarvam coverage: Sarvam Bulbul V4 voice and the Sarvam code agent benchmarks. For another take on judging models and their behavior, see OpenAI's eval-production gap research.
Related reading
- Sarvam AI models and API guide
- Sarvam Epoch 2026 announcements
- Sarvam Bulbul V4 text to speech
- Sarvam code agent benchmarks
- Hugging Face SafeAgents local proxy
- OpenAI eval-production behavior gap
Figures reflect Sarvam's November 3, 2026 launch post and may change.