Sarvam Prism is a new kind of AI model that does not write essays, poems, or code. It makes decisions — and it makes them in 44 milliseconds. Launched on November 3, 2026 by Bengaluru-based Sarvam AI, Prism is built for the thousands of small, fast choices that keep a business running: routing a support ticket to the right team, flagging a suspicious payment, checking whether an AI agent actually finished its job, or picking the best answer from a short list. Instead of generating text one token at a time like a chatbot, Prism reads the situation, looks at the options, and returns a probability for each one in a single pass.
That design choice is what makes it so fast, and it is also what makes it interesting. In this guide, we explain what Prism is, how it works under the hood, how it scored against rival decision models, what it costs, and who should actually use it.
Table of Contents
- Sarvam Prism at a Glance
- What Is Sarvam Prism?
- System One Thinking: Why Fast Decisions Need a Different Model
- How Sarvam Prism Works
- Sarvam Prism Benchmarks: 71.8 at 44 Milliseconds
- Calibration You Can Trust
- Judging AI Agents: Sarvam Prism vs Claude Haiku 5.5
- Security Guardrails: Catching Phishing and Jailbreaks
- Sarvam Prism vs the Competition
- Pricing and Availability
- Who Should Use Sarvam Prism?
- How to Try Sarvam Prism
- Limitations and Honest Caveats
- The Bigger Picture: The Rise of Decision Models
- Frequently Asked Questions
- The Verdict
Sarvam Prism at a Glance
| Detail | Sarvam Prism |
|---|---|
| Developer | Sarvam AI (Bengaluru, India) |
| Launched | November 3, 2026 |
| What it is | Open-weight decision model — picks the best option from a list instead of generating text |
| Base model | Sarvam 105B (105 billion parameters, LoRA fine-tuned) |
| Average benchmark score | 71.8 across three decision indices |
| Response time | 44 milliseconds (median, one NVIDIA B200 GPU) |
| Context window | 131,072 tokens per question |
| Licence | Apache 2.0 (open weights on Hugging Face) |
| Hardware needed | One GPU with at least 120 GB of memory (FP8) |
| Hosted access | Sarvam API on Sarvam Indus |
| Best for | Support routing, agent evaluation, document checks, security screening |

What Is Sarvam Prism?
Sarvam Prism is an enterprise decision model. That phrase deserves a careful definition, because it describes a different job from the one most AI models do. A chatbot answers an open question with free-form text. A decision model does something narrower and much more common in real work: it takes a situation, a question, and a fixed list of options, then tells you how likely each option is to be correct. Think of the difference between asking a colleague to write a report and asking the same colleague to pick the right department for a customer complaint. Both are useful, but the second happens far more often — and it needs to happen fast.
Most of the work inside a company is a long chain of such small choices. A support ticket arrives: is it a billing problem, a technical fault, or a cancellation request? A payment comes in: does it look legitimate or suspicious? An AI agent finishes a task: did it succeed, and does a human need to check it? Each of these is a decision with a short list of possible answers. Prism is designed to handle exactly this kind of work at machine speed, with a confidence score attached to every answer so that a business can automate the easy cases and send only the uncertain ones to a person.
The model is built on Sarvam 105B, the company’s 105-billion-parameter foundation model, and it was adapted for decision-making with a technique called LoRA fine-tuning. LoRA, short for low-rank adaptation, adjusts a small set of extra weights instead of retraining the whole model — a practical way to teach a large model a new skill without starting from scratch. The result is a model that keeps the broad knowledge of its large base while behaving like a specialist judge rather than a general writer.
One of the most important facts about Sarvam Prism is that its weights are open. Sarvam released them on Hugging Face under the Apache 2.0 licence, which means any enterprise can download the model, fine-tune it on its own data, and run it on its own infrastructure. That matters for companies that handle sensitive information — banks, hospitals, insurers, government agencies — because they do not have to send their data to somebody else’s servers to use the model. It also matters for cost: once the model is running on your own hardware, there is no per-request bill from an API provider.
Prism arrives at a moment when the AI industry is discovering that the expensive part of using AI agents is not generating text, but checking the work. Companies now run fleets of agents that browse websites, answer customers, and process documents, and every one of those runs needs a verdict: did it work, or did it go wrong somewhere? A fast, cheap judge is the missing piece in that pipeline, and that is the gap Prism is trying to fill.
System One Thinking: Why Fast Decisions Need a Different Model
To understand why Sarvam Prism exists, it helps to borrow an idea from psychology. In his book Thinking, Fast and Slow, Daniel Kahneman described two modes of human thought. System One is fast, automatic, and intuitive — the kind of thinking you use when you recognize a face or sense that a sentence sounds wrong. System Two is slow, deliberate, and analytical — the kind you use when you solve a maths problem or weigh a major purchase. Most of a working day runs on System One: quick judgements, made in a fraction of a second, that are usually right.
Sarvam AI argues that enterprise work is the same. Much of what a business does in a day is a sequence of fast decisions, and a well-calibrated decision model fits neatly into those workflows. The company did not invent the category — TypeSafe created it last month with its Jev model, named directly after Kahneman’s System One — but Prism is the most ambitious entry in it so far, because it brings a 105-billion-parameter base model to a category that has mostly been served by much smaller, narrower systems.
The practical difference shows up in the numbers. A general-purpose language model asked to judge a decision will typically write out its reasoning token by token, which takes seconds and costs real money at scale. Sarvam Prism skips the essay entirely. It reads the situation once and returns a probability for each option. That is why it can answer in 44 milliseconds while a chat model needs several seconds for the same judgement — and why it can be dramatically cheaper when a company needs millions of these judgements a day.
There is also a reliability argument. When a model writes a long explanation, it can drift, contradict itself, or bury the actual answer in the middle of a paragraph. A decision model returns a structured result: option A, 82 percent; option B, 12 percent; option C, 6 percent. That format is easy for software to act on. A support system can route the ticket automatically when confidence is above 90 percent and send it to a human when it is not. The confidence number is not decoration — it is the mechanism that makes automation safe.
How Sarvam Prism Works
The core idea behind Prism is simple to state and clever in execution: one forward pass, no generation. A normal language model produces its answer by predicting one word, then the next, then the next — a process that repeats dozens or hundreds of times for a single answer. Prism does the whole job in one pass through the network. It takes three inputs — the state (the background information), the question, and the list of options — and a small scoring head at the end returns a probability for each option. There is no token-by-token generation at all.
How Sarvam Prism Scores Options in a Single Pass
Inside the model, the state and the question form one block that is read in both directions, so every word in that block can see every other word. Each option, by contrast, can see the state-and-question block and itself, but never the other options, and every option starts from the same position. Only the final answer token sees everything, and that is where the options are compared. This isolation matters: it means each option is judged on its own merits rather than being influenced by the options around it. According to Sarvam, reading the block in both directions adds about three points on the Jev Decision Index compared with the usual left-to-right reading.
The engineering behind the speed is equally deliberate. Sarvam Prism started at 664 milliseconds per request, running in 16-bit precision across two B200 GPUs. The team then converted the weights to FP8 — a compact 8-bit format — so the model fits on a single 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 fine-tuning adapter into the model weights. Together, these steps brought the median request down to 44 milliseconds, roughly fifteen times faster than where the project started, with no drop in accuracy. The FP8 conversion itself changed the Jev Decision Index score by only 0.01 points, which is effectively nothing.
Training also explains the model’s most unusual strength: its calibration. Prism learns a probability distribution over the options through two equally weighted training objectives. The first is cross-entropy with 5 percent label smoothing, which discourages complete certainty — the model is trained to leave a little room for doubt. The second is a Brier loss, which measures squared probability error and makes confident mistakes expensive. Sarvam credits these two choices for the model’s well-calibrated confidence scores, and the test results, which we cover below, back that claim up.

Sarvam Prism Benchmarks: 71.8 at 44 Milliseconds
Sarvam tested Sarvam Prism against 24 other decision models across three indices, timing every model it could run on the same hardware — one NVIDIA B200 GPU, using each developer’s own inference code. The headline result is a 71.8 average score at a 44-millisecond median response time. That makes Prism the highest-scoring model among all models that answer in under 100 milliseconds. A few models score higher overall, but they take between two and six times as long to respond, which defeats the purpose of a fast decision layer.
| Model | Jev Decision Index | Artificial Analysis (recreated) | Indian languages | Average | Median time |
|---|---|---|---|---|---|
| Sarvam Prism | 67.0 | 65.9 | 82.5 | 71.8 | 44 ms |
| Perplexity Decider v1.2 | 67.7 | 61.2 | 75.0 | 68.0 | 117 ms |
| Jev 1.14 | 60.7 | 61.9 | 70.8 | 64.5 | 196 ms* |
| Clef | 61.7 | 54.8 | 71.2 | 62.6 | 222 ms |
| Microsoft-Decision-1 | 57.8 | 55.6 | 70.0 | 61.1 | 85 ms* |
* Jev 1.14 and Microsoft-Decision-1 are available only through an API, so Sarvam plots their publicly reported times. Measured from Sarvam’s own cluster, they took 431 ms and 533 ms respectively — a reminder that API latency and raw model speed are not the same thing.
The three indices tell different stories, and all three matter. The Jev Decision Index covers 37 public benchmarks. The Artificial Analysis recreation replaces a private knowledge test with curated hard questions that include an “I don’t know” option — and this is where Prism shines most, because its calibrated decisions reward it for admitting ignorance instead of guessing. The third index is an internal Indian-language benchmark, where Prism scored 82.5, well ahead of every rival. That last result is no accident: Sarvam built its reputation on Indian languages, and a decision model that works in ten Indian languages plus Hinglish has a genuine home-market advantage.
One detail in the methodology deserves praise. Sarvam timed each model on the same hardware, sent 1,000 requests one at a time, and reported the median — not a cherry-picked best case. It also published the caveat that API-only models were timed differently. These are the marks of a benchmark report that wants to be checked, and they make the 71.8-at-44-ms headline more believable than the average vendor claim. That said, the benchmarks are still Sarvam’s own tests, not independent verification, and readers should treat them as company-reported until outside labs reproduce them.
Calibration You Can Trust
Speed and accuracy mean little if a model’s confidence scores are untrustworthy. The whole point of automating decisions is that a business acts on the model’s answer only when confidence is above a threshold — say, 90 percent. That system is safe only if the confidence number means what it says: when the model claims 90 percent confidence, it should be wrong roughly 10 percent of the time, not 30 percent.
Sarvam tested this directly. On English questions answered with at least 90 percent confidence, Sarvam Prism predicted a 3.3 percent error rate and actually made errors 2.2 percent of the time — it was, in other words, slightly more cautious than its own confidence suggested. Its closest rival, Clef, predicted 2.8 percent and made 2.3 percent. The other models were far worse: Perplexity Decider underestimated its errors by a factor of 1.5, Microsoft-Decision-1 by 1.8, and Jev 1.14 by 2.7. In Indian languages the gap was even wider: Prism predicted 5.4 percent errors and made 4.4 percent, while rivals made between 2.7 and 4.8 times as many errors as their confidence scores predicted.
The practical consequence is straightforward. With a poorly calibrated model, a business that automates everything above 90 percent confidence is unknowingly automating a large number of mistakes. With a well-calibrated model, the threshold actually works as intended: the automated cases are safe, and the uncertain cases go to a human. Sarvam reports that Prism makes fewer errors than its confidence scores predict at every threshold tested from 50 to 95 percent, in both English and Indian languages. If that holds up in independent testing, calibration — not raw accuracy — may be Prism’s most valuable feature.
There is a second consistency test that matters just as much. Changing the order of options can change a model’s answer, which is a serious problem when a decision affects a customer. Sarvam ran 2,000 questions four times with shuffled options and four times without shuffling. Rivals changed between 4 and 12 percent of their answers when the options were shuffled. Sarvam Prism, after a fix that puts options into one fixed order before scoring and returns each probability under its original name, changed zero of 2,000 answers. For regulated industries, where the same case must get the same verdict regardless of presentation order, that kind of determinism is not a nicety — it is a requirement.
Judging AI Agents: Sarvam Prism vs Claude Haiku 5.5
Perhaps the most commercially important test in Sarvam’s announcement has nothing to do with customer support at all. It is about judging AI agents. As companies deploy more agents, they need a reliable way to answer three questions about every run: did the task succeed, does a human need to review it, and — when something failed — which agent caused the failure? Answering those questions with a full-size chat model is slow and expensive, because the model must read a long trace and write out a verdict.
Prism scored a full agent run in about half a second on a single GPU, reading the trace once and answering all five evaluation questions together. The test used 700 recorded agent runs and 1,770 questions drawn from AgentRewardBench (web-agent runs labelled by experts), τ-bench (customer-service conversations labelled by task completion), and Who&When plus TRAIL (failed runs labelled with the responsible step and agent). Every answer was checked against a known correct label, never against another AI model’s judgement.
The results put Sarvam Prism ahead of every rival on all three questions, including Claude Haiku 5.5, a well-regarded small model from Anthropic that many teams currently use as a judge. Prism scored 75.0 percent on whether the task succeeded (Haiku 5.5: 66.0 percent), 86.4 percent on whether a run needs human review (Haiku 5.5: 77.9 percent), and 69.6 percent on which agent caused a failure (Haiku 5.5: 37.0 percent). Overall, across all questions, Prism reached 78.5 percent against Haiku 5.5’s 70.7 percent.
The cost comparison is just as striking. Claude Haiku 5.5, called through its API, takes a median 2.5 seconds per run and costs $0.64 per 1,000 runs.
Prism does the same job in about half a second on one GPU, with no per-run API bill once the model is deployed. For a company evaluating millions of agent runs a month, that difference is the gap between a quality-control system it can afford and one it cannot. It also explains why Prism matters beyond India: the agent-evaluation problem is global, and a fast, open-weight judge is useful everywhere. Readers interested in the agent wave should also see our coverage of DeepSeek’s super agent and Claude’s message threads, which show how fast this market is moving.

Security Guardrails: Catching Phishing and Jailbreaks
The second specialist test covers security. Sarvam built an internal benchmark of 5,517 questions across nine tasks, drawn from vulnerabilities published since July 2026. The model is asked to classify whether an artifact is part of a phishing campaign, whether an incoming prompt is trying to jailbreak the system, and what type of attack is involved, mapped to MITRE frameworks — the standard vocabulary security teams use to describe adversary tactics.
Sarvam Prism scored 89.0 percent on this benchmark, ahead of Clef at 84.5 percent, Jev 1.14 at 78.6 percent, and Microsoft-Decision-1 at 76.1 percent.
The company says it will expand this evaluation set and release it publicly in the coming weeks, which is the right move: an internal security benchmark that nobody else can inspect is a claim, not evidence. Until that release happens, the 89.0 percent figure should be read as promising but unverified. Still, the use case is easy to grasp. A security filter that can classify a suspicious message in 44 milliseconds can sit in front of every AI interaction in a company, catching prompt-injection attempts and phishing content before they reach a more expensive model or a human employee.

Sarvam Prism vs the Competition
The decision-model category is only weeks old, but it already has a clear field. Here is how the main contenders compare on Sarvam’s own measurements:
| Model | Maker | Average score | Speed | Availability |
|---|---|---|---|---|
| Sarvam Prism | Sarvam AI | 71.8 | 44 ms | Open weights (Apache 2.0) |
| Perplexity Decider v1.2 | Perplexity | 68.0 | 117 ms | API |
| Jev 1.14 | TypeSafe | 64.5 | 196 ms (reported) | API |
| Clef | — | 62.6 | 222 ms | Weights |
| Microsoft-Decision-1 | Microsoft | 61.1 | 85 ms (reported) | API |
The pattern is clear. Prism leads on accuracy and on speed at the same time, which is unusual — normally you trade one for the other. Its open weights are the other differentiator: Perplexity Decider, Jev, and Microsoft-Decision-1 are API-only, which means your data leaves your building and you pay per request. Prism’s main weakness is its hardware requirement. A GPU with 120 GB of memory is not a laptop accessory; it is data-center equipment, and small companies will realistically use the hosted Sarvam API rather than self-hosting. The honest comparison, then, is not “free versus paid” but “self-hosted control versus API convenience” — and different companies will choose differently.
It is also worth noting what Sarvam Prism is not competing with. It is not a rival to ChatGPT, Claude, or Gemini, and it will not write your emails. The right comparison is with the judging and routing layers that sit around those models. If you already use a frontier model for customer-facing work, Prism is a candidate for the layer that checks, routes, and supervises that work — not a replacement for the model itself.
Pricing and Availability
Prism is available right now in two forms. The weights are on Hugging Face under the Apache 2.0 licence, which permits commercial use with almost no restrictions — a company can fine-tune the model on its own data and run it on its own infrastructure without paying Sarvam a licence fee. The only cost is hardware: one GPU with at least 120 GB of memory, running the model in FP8. That is a serious but standard server GPU, and for a company already running AI infrastructure it is a modest addition.
The second route is the hosted Sarvam API on Sarvam Indus, which lets teams use Sarvam Prism without buying hardware at all. Sarvam has not published a separate per-request price for Prism at the time of writing, so teams evaluating the hosted route should check the Sarvam Indus pricing page before committing. The open-weights route, by contrast, has a clear cost structure: hardware plus engineering time, and no per-decision meter running. For high-volume use — millions of decisions a day — self-hosting will almost always win on cost. For a pilot project or a small team, the API is the sensible starting point.
Context matters here. The decision-model category is so new that pricing norms have not settled yet. API-only rivals charge per request, which is simple but punishes exactly the high-volume use that decision models are built for. Sarvam’s open-weights play is a bet that enterprises would rather own the layer that judges their AI than rent it by the judgement — and given how sensitive the data flowing through a decision layer can be, that bet looks reasonable. For readers tracking the open-weights trend, our Tencent Hy4 guide and Drex Max open-weights guide show the same pattern playing out across the industry.
Who Should Use Sarvam Prism?
Prism is not a consumer product, and ordinary users will probably never interact with it directly. Its audience is engineering and operations teams, and the use cases Sarvam has already tested point to four groups that should pay attention.
- Customer support teams. Intent routing is the classic decision-model job: read a message, pick the right category, and send it to the right team or automated flow. Sarvam Prism’s speed means routing adds no perceptible delay to a conversation, and its calibration means low-confidence cases can be escalated to a human instead of being misrouted.
- Companies running AI agents. If you deploy agents that browse, book, or process documents, you need a judge. Prism’s agent-trace scores — especially its lead on identifying which agent caused a failure — make it a strong candidate for the quality-control layer of an agent system.
- Security and trust-and-safety teams. Phishing classification, jailbreak detection, and attack-type labelling are all decisions with short option lists. A 44-millisecond filter can sit in front of every AI interaction without slowing the product down.
- Document-heavy businesses. Document intelligence — classifying, extracting, and checking long documents — benefits from Sarvam Prism’s 131,072-token context window, which lets it evaluate long agent traces and long documents in a single pass rather than splitting them into chunks.
There is also a geography angle. Prism’s Indian-language scores (82.5, far ahead of rivals) make it the strongest decision model currently available for businesses serving customers in India, where support conversations routinely mix Hindi, Hinglish, and English. For US readers, the English-language scores are the relevant ones — and there Prism still leads, just by a smaller margin.

How to Try Sarvam Prism
Trying Sarvam Prism depends on which route fits your team. The fastest way to experiment is the hosted Sarvam API on Sarvam Indus: sign up, send a request containing a state, a question, and a list of options, and read back the probability for each option. That is enough to test routing or classification on your own data within an afternoon.
Teams that want to self-host should start from the Hugging Face weights. You will need a GPU with at least 120 GB of memory and the FP8 inference setup Sarvam documents. The workflow is the same — state, question, options in; probabilities out — but the model runs inside your own environment, which is the right choice for regulated data or very high volumes. Sarvam’s documentation covers the request format and the confidence-threshold pattern: act automatically above your chosen threshold, and route everything below it to a human for review.
A sensible pilot looks like this. Pick one decision your team already makes thousands of times a week — support routing is the easiest. Run Prism alongside your current system for two weeks, compare its picks against your existing labels, and check two numbers: accuracy and calibration. Accuracy tells you whether it picks the right option. Calibration tells you whether its confidence scores can be trusted for automation. Only automate the cases above a threshold you have verified on your own data, and keep measuring after launch, because real-world data drifts and yesterday’s calibration is not tomorrow’s guarantee.
Limitations and Honest Caveats
No new model deserves uncritical coverage, and Sarvam Prism has real limitations that buyers should weigh. First, every benchmark in this article comes from Sarvam itself. The methodology looks careful — same hardware, median over 1,000 requests, published caveats — but independent labs have not yet reproduced the results, and the security benchmark is still internal, with a public release promised in the coming weeks. Treat the numbers as the company’s best case until outsiders confirm them.
Second, the hardware bar is high. A 120 GB GPU excludes small businesses and individual developers from self-hosting, which means the open-weights advantage is mostly an enterprise advantage. Smaller teams will use the hosted API, where Sarvam has not yet published Prism-specific pricing — so the total cost of that route is still unclear.
Third, Prism answers questions with fixed options. It cannot handle open-ended requests, and it cannot explain its reasoning the way a chat model can. That is a design choice, not a flaw, but it means teams still need a general model alongside it for anything that is not a decision. Finally, the Indian-language advantage, while real, is Sarvam’s home turf; US buyers should judge the model on its English scores, where its lead over Perplexity Decider is narrower (67.0 versus 67.7 on the Jev index, though Prism leads on the overall average and on speed).
The Bigger Picture: The Rise of Decision Models
Sarvam Prism is best understood as part of a shift in how AI gets used inside companies. The first wave of enterprise AI was about generation: write the email, summarize the document, answer the question. The second wave, now arriving, is about judgement: check the work, route the request, flag the risk, pick the option. Generation gets the headlines, but judgement is where most of the volume is — and volume is where speed and cost decide everything.
That is why a 44-millisecond model matters more than it might seem. When a judgement costs almost nothing and takes almost no time, it can be applied everywhere: to every support message, every agent run, every suspicious input. Quality control stops being a sampling exercise — checking one run in a hundred — and becomes universal. The companies that adopt decision models early will have a structural advantage: their agents will be checked, their support will be routed correctly, and their security filters will run on every interaction rather than a sample.
The category is barely a month old — TypeSafe named it with Jev in October 2026, and Prism is its most capable entry yet. Expect rapid change: more entrants, independent benchmarks, and pricing pressure as open weights force API providers to compete. For now, Prism sets the bar: 71.8 average, 44 milliseconds, open weights, and calibration that actually holds up. Rivals will have to beat all four, not just one.
Frequently Asked Questions
What is Sarvam Prism used for?
Sarvam Prism is used for fast enterprise decisions with fixed options: routing support tickets, evaluating AI agent runs, screening content for phishing and jailbreak attempts, and classifying documents. It returns a probability for each option instead of generating text.
How fast is Sarvam Prism?
Prism answers in a median 44 milliseconds on a single NVIDIA B200 GPU, according to Sarvam’s tests. That is roughly fifteen times faster than the 664-millisecond starting point of the same model before its FP8 and kernel optimizations.
Is Sarvam Prism free?
The weights are free under the Apache 2.0 licence — you can download, fine-tune, and run them commercially. You still pay for the hardware (one GPU with at least 120 GB of memory) or for hosted access through the Sarvam API on Sarvam Indus.
How does Sarvam Prism compare with Perplexity Decider?
On Sarvam’s tests, Sarvam Prism scores higher overall (71.8 versus 68.0) and answers nearly three times faster (44 ms versus 117 ms). Perplexity Decider edges ahead on one index, the Jev Decision Index (67.7 versus 67.0), but trails on the other two indices and on calibration.
Can Sarvam Prism replace ChatGPT or Claude?
No. Prism does not generate text and cannot answer open-ended questions. It is a judging and routing layer that works alongside general models, checking their work and making fast structured decisions.
The Verdict
Sarvam Prism is the most interesting enterprise AI launch of the week, not because it is the biggest model, but because it solves the problem companies actually have right now: judging AI work quickly, cheaply, and reliably. Its headline numbers — 71.8 average score, 44-millisecond responses, open weights under Apache 2.0 — are genuinely impressive, and its calibration results may matter even more than its accuracy, because calibration is what makes automation safe. The agent-evaluation scores, where it beats Claude Haiku 5.5 on every question while costing a fraction as much per run, point to its most valuable role: the quality-control layer for the age of AI agents.
The caveats are real but manageable. Benchmarks are company-reported, the security test set is not yet public, and self-hosting needs serious hardware. None of those should stop a team from running a pilot — they should simply shape how the pilot is measured. If you run support, agents, or security screening at scale, Prism deserves a two-week trial on your own data. If it holds up there, it is not just a faster model; it is a cheaper, more reliable way to run the judgement layer of your whole AI operation. For the official announcement, see Sarvam’s Prism launch post, and for independent coverage, see Unite.AI’s report on the release.
