Sarvam AI just open sourced a model that skips writing sentences entirely and spits out a probability in 44 milliseconds, built specifically for the moment an AI agent has to pick one option and move on.
Most AI headlines this year have been about models getting smarter. Sarvam's newest release bets on something narrower: getting fast enough that speed itself becomes the pitch. The Bengaluru startup released Prism, an open-weight decision model that returns a calibrated probability for a set of answer options without generating a single word of text. According to AlphaSignal's coverage of the release, Prism is a LoRA fine-tune built on top of Sarvam's own 105-billion-parameter mixture-of-experts model. It comes under an Apache 2.0 license and is available both on Hugging Face and through Sarvam's hosted Indus token factory.
The trick is architectural, not just a smaller model crammed onto faster hardware. A normal language model, even a fast one, generates text token by token, scoring tens of thousands of candidate words at every step. Prism throws that away. It computes scores directly from the model's hidden states. Those feed into a small decision head that converts them into one probability per option, whether that's "fraud" versus "not fraud," a routing choice between five possible agents, or a confidence score on a scale. No sentence gets written, so there's nothing to wait on.
That distinction matters because the market Sarvam is chasing isn't chatbots. It's the backend of agentic AI systems, the orchestration layer that decides which sub-agent handles a task, whether a transaction looks suspicious, or which response variant a customer-service bot should route to. Sarvam says it already uses Prism internally for trace evaluation, prompt guardrails, routing decisions inside its token factory, and judging outputs in its speech-to-speech testing harness. Those are exactly the places where a 500-millisecond delay, trivial in a chat interface, becomes a bottleneck when it happens hundreds of times inside a single automated workflow.
On benchmarks, Prism scored an average of 71.8 across three decision-accuracy indices, which Sarvam's own reporting frames as the best result among sub-100-millisecond models, per AlphaSignal. Sarvam also disclosed something less flattering, and more credible for it: Prism runs underconfident across every confidence band it was tested on, in both English and Indian languages. That means its actual error rates came in lower than its stated probabilities implied. It's a specific, checkable claim in an industry where vague "state-of-the-art" talk is the norm. And underconfidence is a far safer failure mode for a fraud filter than a model that's more sure of itself than it should be.
Sarvam AI's New Coding Agent Turns Out To Be A Rebadged OpenAI Codex Sarvam AI unveiled Sarvam Code at its Epoch 2026 conference in Bengaluru, pitching it as a cheaper, India-hosted rival to Claude Code and OpenAI's Codex. A technical teardown of the 348MB binary found it shares Codex's exact commands, config schema and an unscrubbed 'x-openai-subagent' string, raising attribution questions under Codex's Apache 2.0... - sarvam AI coding agent rebadged OpenAI Codex - India built coding tool actually OpenAI Codex
Context helps here too. Groq and Cerebras have spent years building custom silicon, Groq's LPUs and Cerebras's wafer-scale chips, to push large language model inference latency down to roughly 80 to 150 milliseconds time-to-first-token. That's on specialized hardware with a narrow catalog of ported models. Prism is getting under half that figure without needing bespoke chips at all. It gets there by not generating text in the first place, and by shrinking the computation to only the handful of tokens an answer actually requires rather than scoring the model's full vocabulary.
There's a catch worth being straight about. Prism's custom FP8 layout isn't compatible with standard tools like AutoModelForCausalLM or general-purpose inference servers, according to AlphaSignal's reporting. That means teams either run it through Sarvam's own serving stack or do nontrivial integration work to self-host it. It's the kind of detail a company announcing a victory lap usually leaves out, and it's the kind of detail that tells you whether the claim is real.
Sarvam's bigger play here is sovereignty, not just speed. The company has raised roughly $300 million at a valuation of about $1.5 billion, according to LetsDataScience, with later reports suggesting a fresh round could push that figure toward $3 billion. It has positioned itself as the anchor of India's push for homegrown AI infrastructure, building everything from its 105B foundation model to the Indus agent platform that bundles coding, document, and work agents under one roof. Prism fits that stack as the piece meant to make all those agents cheap to run at scale, rather than the piece meant to impress anyone with raw intelligence.
Frankly, that's a smarter bet than chasing frontier scale against OpenAI or Anthropic. Decision latency is a problem every company running agents at volume actually has, today, and it's one Sarvam can credibly claim to have solved with a number anyone can go test for themselves.
Also read: Lovable AI Apps Left Passport and Payment Data Open to Anyone • Revolut Shows Cheap Unlimited Tokens, Not Smarter AI, Is What Speeds Up Coding • Mistral Is Racing OpenAI and xAI on a Fraction of Their GPUs
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.