Two things happened to the same Indian AI company this week, and they pull in different directions. Sarvam, the Bengaluru lab that has spent the last two years building its own Indian-language models, shipped a genuinely useful one: a speech-to-text system that understands all twenty-two official Indian languages, switching between them automatically without being told which one is coming. Two days later, the same company said it would also start hosting other companies' AI models on its own Indian servers, so a business here could rent an American model without the request ever leaving the country. Those are two different claims — "we built this" and "we host this" — and they are easy to blur together. Meanwhile Mozilla put a French AI company's model inside Firefox's new browsing assistant, routed through a proxy that strips out anything that could identify who is asking. And the cheapest way to make an AI system answer for less money this week did not come from a bigger model. It came from teaching an existing one to think in fewer words.
Why we build it this way
We read a story like Sarvam's for one reason: buyers are starting to ask the sovereignty question, and most vendors can answer it in a sentence that sounds reassuring whether or not it is true. "Hosted in India" and "never leaves your building" sound similar and are not. A model hosted in an Indian data centre is still a model somebody else runs, on somebody else's schedule, with a support ticket standing between a question and its answer. A model that runs entirely on hardware the customer owns has no such gap: there is nobody to ask permission, and nobody who can turn it off from outside.
"Hosted in India" and "runs entirely on hardware you own" are two different claims. A buyer who cannot tell them apart will pay for the first while believing they bought the second.
Sarvam becoming a host for other companies' models is not a bad idea. It is probably good business, and it will genuinely help Indian companies that want an American model without sending their data abroad. But it changes what the word "sovereign" means when Sarvam says it, and it is worth us being precise about which version we are. We do not host models. We hand a customer a system that runs on a box inside their own building, with nothing to renew and no one else's servers in the loop.
The Firefox story makes the same point from the other direction. Mozilla did not solve privacy by keeping its assistant on the user's own machine — Firefox's assistant still sends the question to a server. What it did instead was strip out everything that could identify the sender, and get its AI partner to promise not to keep a copy. That is a real improvement, and worth taking seriously. But it is still a promise a company makes about your data, not a fact about where your data physically sits. That difference is the whole of what we build around.
What changed this week
- Sarvam released Saaras V4, a speech-to-text model that understands all 22 official Indian languages plus English, switching between them mid-sentence. Confirmed by Sarvam's own announcement, 25 September 2026.
- Sarvam also said it will host other companies' AI models, including American ones, on its own Indian infrastructure, not only its own models. Said by Sarvam's co-founder at a public conference and reported by several outlets, 26 September 2026 — not yet in a written company statement, so we count it as the company's own claim rather than a confirmed policy.
- Firefox's new browsing assistant, Smart Window, is now powered in part by a Mistral AI model. Every request is routed through a Mozilla-run proxy that strips out anything that could identify the person asking, and Mistral has agreed not to keep the conversations. Confirmed by both companies, 16 September 2026.
New models, and what they now cost
Alibaba's Qwen team released a model that reads and writes text and also understands pictures, audio and video in a single system, aimed at tasks like listening to a call and then acting on what was said. Audio in particular got sharply cheaper to process. Confirmed by Qwen's own release, 18 September 2026. Separately, at a conference the same week, Alibaba said its next major model, Qwen 4, is already being trained and will arrive "soon" — with no date, price or way to run it yet, so treat that as a plan rather than a product.
What people are actually trying this week
Fireworks AI, a smaller American AI company, took an existing large open model and retrained it to reach the same answers while "thinking out loud" in far fewer words — which matters because most AI systems charge by the word, including the words spent thinking before the final answer arrives. Fireworks published its own test results and says the savings hold up without a drop in quality. Nobody outside the company has checked that independently yet, so it is worth watching rather than relying on. 23 September 2026.
Separately, the electronics maker Xiaomi gave away, for free, one of the largest AI models released so far this year, along with the full recipe for how it was trained — the code, the training setup, and a technical report explaining the choices. An independent research group that tracks freely downloadable AI models placed it at the top of its own leaderboard within days. Nobody builds a product on a model this new without testing it themselves first, but it is a genuine data point that a very large, very capable model can be trained cheaply and given away complete. 21 September 2026.
Explained simply: reasoning tokens
Before some AI systems give a final answer, they write out a private chain of working-through-it text that you do not usually see — and you are billed for every word of it, the same as for the answer itself.
Picture a shop assistant you pay by the minute. A confident one glances at the shelf and answers in five seconds. A careful one paces the aisle, thinking out loud, for two minutes before giving the same answer. Both get you the right size. One costs more to ask.
Letting an AI system "think" for longer, in more of these hidden words, often makes it more accurate on hard problems — that is one of the real advances of the last two years, and several labs have shown it independently. But it also means two AI companies can quote very different prices for what looks like the same task, and the difference is often just how many hidden words one of them used to get there.
That thinking longer can improve accuracy is well established. What Fireworks claimed this week with its retrained model — that you can cut those hidden words sharply without losing the accuracy they bought you — is a newer, single-company claim. It is plausible, since other labs have shown smaller versions of the same effect, but nobody outside Fireworks has reproduced its specific numbers yet.
Nothing we build is billed by the word — our systems run on hardware the customer already owns, so there is no per-word invoice to shrink. But it is worth knowing the shape of this argument, because a customer comparing two cloud AI quotes will eventually ask why one costs several times the other for what sounds like the same job. Often the honest answer is exactly this: one model spent far more hidden words thinking about it than the other did.
None of this changes what we ship this week. What it sharpened is a distinction worth making out loud every time it comes up: a model hosted in India is not the same claim as a model that runs entirely inside a customer's own building, and a company's promise not to keep your data is not the same as your data never leaving the room. We will keep reading the same sources next week, and if it is a quiet one, we will say so instead of padding it.
Read about VarahiOne →