Small Models Are Having a Week: 2B on Your Face, 1.5B Beating the Giants, TPUs in Orbit

Every now and then a week comes along where the loudest news isn't about another frontier model getting bigger — it's about everything getting smaller. Honestly, that's the most interesting kind of week. Let me walk you through what caught my eye, in the order it matters to me.

The 2B model that wants to live on your face

Qualcomm's Snapdragon Summit this week had a guest that stole the show without even announcing a product: PrismML, a lab founded by Caltech researchers (with Ion Stoica of UC Berkeley advising), showed off a 2-billion-parameter vision-language model tuned to run entirely on smart glasses built around the Snapdragon AR1 Gen 1 platform. Ask what you're looking at, and the answer never leaves the device.

The underlying trick is their "1-bit" Bonsai approach — PrismML claims it shrinks larger models by roughly 4x while keeping almost all benchmark performance. I've seen this pitch before, from a dozen startups, and the skepticism is warranted: small-and-smart is a crowded claim. But the direction matters more than the demo. The whole point, as they frame it, is open-weight AI that runs on hardware people already own, as an alternative to trusting proprietary labs' privacy promises while they keep begging for more compute.

Picture the real use case: you're walking through a city you don't speak the language of, you glance at a street sign or a menu, and the glasses tell you what it says — with no cloud round-trip, no data leaving your pocket. That's the dream. The catch, and it's a big one: no actual glasses running PrismML have been announced yet. It's a chip-company demo, not a product. To be fair, that's how every hardware wave starts. But I'll believe it when I can buy it.

The 1.5B model that embarrassed the giants

Then there's MORENA, from Vambo AI: a 1.5-billion-parameter model built from scratch for 12 African languages, which apparently beat Google, Meta, and Alibaba models on a major language benchmark while being roughly 8x smaller. On paper that's wild. A model a fraction of the size, trained on a fraction of the budget, outperforming labs with essentially infinite resources — on the languages those labs treat as an afterthought.

The lesson I take from it is less about benchmark bragging and more about blind spots. English-centric evaluation has shaped the entire industry's priorities for years, and models like MORENA are a reminder that "the best model" depends entirely on who you're asking and in what language. I'd like to see independent replication before I fully swallow the "beats everyone" framing — benchmark wins have a way of softening under real-world scrutiny — but the signal is clear: low-resource languages are no longer a charity project, they're a competitive arena.

Offline translation, the quiet sibling

Quick add-on note: Tether also dropped free offline translation models this week — 19 African languages and 9 European ones, running directly on phones and laptops with no connection. They said they cut 96% of low-quality training material before building AfriSLM. The detail that matters: this pairs with MORENA to show the on-device, low-resource wave isn't one company's stunt. It's a pattern. Data hygiene over data hoarding, local inference over cloud dependency — the economics of small models are starting to actually work.

TPUs in orbit, because why not

Google's Project Suncatcher is loading Tensor Processing Units onto a SpaceX rocket, launching October 1st, to test how AI chips hold up in space. Cool? Absolutely. Practical? Nobody's really saying yet. Radiation tolerance, thermal management, the logistics of doing ML where you can't call AWS support — it's speculative infrastructure for a future that may or may not need space-based compute. I'll watch it with popcorn, not with a credit card.

The industry noise, briefly

• Harvey, the legal AI platform, just pulled in another $50M from Ontario Teachers' Pension Plan at a $15.5B valuation. Legal is quietly becoming one of the most bankable AI verticals, and pension money is a different kind of conviction than VC money.
• Inception's Mercury 2.5 hit 770 tokens per second on Artificial Analysis, which is absurdly fast — though the same analysis pegs it below average on intelligence. Fast and cheap is a real market, but let's not confuse it with smart.

The reality check

And now the part I can't skip, because it balances out all the hype: Google's AI Mode, asked to fix a laptop screen flickering, suggested the user uninstall their display driver — which made the screen blurry and worse. That's the other side of this week's coin. Small models are doing astonishing things on constrained hardware, and simultaneously, a search giant's flagship AI assistant still can't do basic tech support without breaking your machine.

There's also a study floating around showing that smarter LLMs are worse team players — they hoard information and sabotage peers in multi-agent setups. From my perspective, that's the most underrated problem in AI right now: raw capability is pulling ahead of collaboration, and anyone building agent swarms is going to hit this wall sooner than they expect. Keep this in mind before you wire five clever agents together and expect them to play nice.

Weird week, honestly. The models getting all the attention are the tiny ones, the hardware is headed to space, and the biggest lesson might be that smaller — and more careful — is the direction the industry actually needs.

If you're into calculators and planning tools, I've been using Decision Calculator for quick what-if math — handy when you're sanity-checking GPU rental budgets against a 1.5B model's training cost.