Primary use stated or inferred cautiously from official documentation.
S
SARVAM AI / sarvamai/prism · API sarvam-prism
Sarvam Prism
Sarvam AI decision model released on November 3 and based on Sarvam 105B. It processes state, questions and options and returns probabilities in a single pass. It neither generates free text nor accepts images. FP8 weights are published and the API uses sarvam-prism through POST /v2/systemone.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | sarvamai/prism · API sarvam-prism |
|---|---|
| Model type | Classification and decision |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted |
| Release | 03 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · JSON |
| Exit | Probabilities and typed values |
| Context window | API: 32,768 tokens per question; weights: 131,072 |
| maximum output | Probability per option, without text generation |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · PyTorch |
Request classification, moderation, selection among options and assessment of agent runs. Supports choice, yes/no and ordered-scale questions. Confidence must be validated in the real domain; a high probability does not replace available evidence or eliminate bias risk.
The API limits each question to 32,768 tokens, while the weights document 131,072. These channels must not be confused. Local code uses a custom FP8 format, around 106 GB of disk and about 120 GB of free GPU memory. It is not loaded as AutoModelForCausalLM. Speed was validated on B200.
03 / CAPABILITIES
What it can do
Orientation
Request classification, moderation, selection among options and assessment of agent runs. Supports choice, yes/no and ordered-scale questions. Confidence must be validated in the real domain; a high probability does not replace available evidence or eliminate bias risk.
Context
API: 32,768 tokens per question; weights: 131,072 · Probability per option, without text generation
Tools and integration
Not published
Access
Hugging Face and Sarvam API
Input modalities
TextJSON
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | 4 INR / 1M tokens in the API | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | 0 INR for output in the API | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
Sarvam lists 4 INR per million input tokens and 0 INR for output tokens. This is decision API billing, not a price for generated text. Local deployment adds infrastructure and operations. No USD conversion is invented and GPU resources are not assumed to be free.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
The provider reports an average of 71.8 across three decision indices. The card records 44 ms median and 420 ms at the 95th percentile over one thousand sequential requests after loading on a B200. These are provider tests and do not guarantee network latency or performance on other infrastructure.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| Sarvam Prism decision indices average | 71,8 / 100 | Score | View source ↗ |
| Sarvam Prism sequential B200 median latency | 44 ms | Median; lower is better | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Sarvam Prism
Version added to Inferama's verified catalog.
07 / SOURCES