Sarvam Prism brings fast AI decisions to an open model and hosted API
Sarvam Prism launched on November 3, 2026, giving developers a model designed to select an answer from predefined options instead of writing a response. Built on Sarvam 105B, it is available through a hosted API and as downloadable weights. The practical attraction is a decision layer for software that needs a label, a score or a yes/no result, rather than another paragraph to interpret.
That narrower job changes how teams should evaluate it. A customer support queue, for example, needs a dependable assignment to the right team and a sensible route for uncertain cases. Fluent explanations alone do not establish either. Prism makes the probability distribution part of the interface, putting the decision threshold directly in the developer’s hands.
What Sarvam actually measured
In its release report, Sarvam gives Prism an average score of 71.8 across three decision indices and a median response time of 44 milliseconds on one NVIDIA B200. The company says this was the highest average among the tested models responding in under 100 milliseconds. Some competitors scored higher overall.
These are Sarvam’s results, not independent ByteForward tests. The composite includes the public Jev Decision Index, Sarvam’s recreation of an Artificial Analysis index and an internal Indian language benchmark. In that recreation, Sarvam replaces the index’s private knowledge test with curated hard questions. API-only competitors also use reported latency in the comparison, making the chart less uniform than a single hardware benchmark might suggest.
Hosted availability and pricing
Sarvam’s pricing page lists Prism as available at ₹4 per million input tokens and ₹0 for output tokens, excluding GST. The listed token rates apply across plans. Pay as you go has no monthly fee, while verified new users are offered ₹100 in credits.
The shared input state is counted once per request, even when a developer asks multiple questions about it. Question text and answer options still contribute to input usage. For budgeting, this makes the shape of the request worth checking rather than multiplying the full context by every label the application wants to assign.
The API and local model have different limits
The hosted API documentation specifies a 32,768 token limit per question, including context and options. It accepts text or JSON, with no image input or streaming. Choice questions support 2 to 255 options and score questions support 2 to 10 ordered levels. A request must finish within 60 seconds.
The Hugging Face model card instead lists a 131,072 token context limit per question for the downloadable model. It specifies an FP8-capable NVIDIA GPU with about 120 GB of free memory, plus roughly 106 GB of storage for weights. The release carries an Apache 2.0 license.
Self-hosting therefore requires substantial infrastructure. The custom FP8 layout needs Sarvam’s supplied inference code rather than a standard causal language model loader or general-purpose serving stack. The card reports a 420 millisecond 95th-percentile response time alongside the 44 millisecond median, measured over 1,000 sequential requests on a B200. Neither figure establishes hosted end-to-end latency under production traffic.
Read probabilities before automating a decision
One integration detail deserves attention: the hosted choice response includes both probabilities and a rescaled confidence field. Sarvam says its calibration thresholds refer to the top option’s probability, not that confidence field. Treating those two numbers as interchangeable would change which cases pass an automation threshold.
The documentation also reports lower coverage at a 0.9 threshold for Indian languages: about 38.4% of answers clear it, compared with 68.5% in English. More cases would therefore need review. Hosted yes/no questions use the type name noul, while the local example uses yes_no, another reason to follow the documentation for the deployment being used.
The useful next step for an adopting team is a representative evaluation with its own labels, error costs and fallback process. Fixed options remove a formatting problem, but choosing the wrong valid option can still send work to the wrong destination. Measure those errors and the review workload together before treating benchmark confidence as permission to automate a live workflow.
For an audio-focused comparison, see our Hume Audio Decision API coverage, which explains its different probability semantics and waitlist access.
Featured image: conceptual illustration by ByteForward, generated with AI, with the official Sarvam wordmark. Not a product interface or measured result.
