Two things happened in AI this week that read like opposite bets and are actually the same bet, paying off twice. PrismML shrank a genuinely capable reasoning model down to under 4GB and got it running on an iPhone at 90% of its original performance. In the same seven days, Thinking Machines Lab and Moonshot both shipped open-weight models in the trillion-parameter range. Read those as a contest between big labs and small labs and you'll miss what's actually going on: the frontier isn't one race anymore, it's two, and only one of them needs a data center. The same week also handed you a reason to care which side you're on — OpenAI's own agent deleted a user's home directory when given too much rope, which is exactly the kind of mistake that matters less on a machine you control.
The 4GB reasoning model
Start with the smaller number, because it's the more useful one. PrismML's Bonsai 27B is a 1-bit and ternary quantized build of Qwen3.6-27B that runs on laptops and phones. The follow-up detail is the one that matters: it compresses to under 4GB and holds 90% of its original performance running on an iPhone. Ninety percent, under 4GB, on hardware you already own — that's not a demo, that's a deployment target. A year ago "runs on a phone" meant a stripped-down assistant that could summarize a paragraph. This is a reasoning model.
It wasn't a solo release, either. The same week, the Soofi Consortium shipped Soofi S 30B-A3B, an open hybrid Mamba-Transformer mixture-of-experts model for German and English. Different technique — architecture instead of quantization — same instinct: build something right-sized for the hardware people actually have, rather than a smaller cut of something built for a data center. Two labs converging on "efficient by design" in the same seven days isn't a coincidence, it's a direction.
The trillion-parameter counterweight
The same week, the frontier side of the industry did the opposite. Thinking Machines Lab released Inkling, a 975-billion-parameter open-weight multimodal mixture-of-experts model with 41 billion active parameters. Moonshot's Kimi K3 landed at 2.8 trillion parameters, close enough to GPT-5.6 Sol and Claude Fable 5 that the coverage framed it as closing a gap rather than opening one — and the same reporting flagged that the era of dirt-cheap Chinese frontier models might be ending, because models at this size cost real money to train no matter who ships them.
Worth being precise here, because the honest picture has a wrinkle: even the trillion-parameter side isn't brute size all the way down. Inkling's 41 billion active parameters out of 975 billion total is the same mixture-of-experts trick that keeps frontier inference costs from exploding — only a slice of the model fires per token. The difference between the two tracks isn't "sparse versus dense," it's how far each one is willing to take sparsity. Inkling uses it to make a trillion-parameter model affordable to serve. Bonsai uses a more aggressive version of the same instinct — ternary and 1-bit weights — to make a 27-billion-parameter model fit somewhere it has no business fitting. Picture the industry right now less as one race up a single mountain and more as two separate expeditions that happened to start from the same base camp — one is climbing, the other is heading downhill toward you, and both are traveling as light as the engineering allows.
What brute force buys you
The scale side isn't wasted effort, to be fair — it buys real capability. GPT-5.6 Sol Ultra reportedly proved the fifty-year-old Cycle Double Cover Conjecture in under an hour by running 64 subagents in parallel, and a University of Pennsylvania professor used GPT-5.6 Sol to disprove a thirty-year-old statistics conjecture in ninety minutes, after humans hadn't managed it. That's what a trillion-plus parameters and a warehouse of compute gets you: brute-force throughput on problems nobody has cracked, run by throwing 64 parallel agents at a proof instead of one careful one. Genuinely impressive. Also genuinely not what a small business or an individual practitioner needs on a Tuesday — a law firm drafting contracts or a clinic doing patient intake isn't short on mathematical horsepower, it's short on a model it can trust with the data in the first place.
Here's the part that should make you careful about which side you pick, and it landed the same week as everything above: GPT-5.6 accidentally wiped users' home directories when running in Full Access Mode. OpenAI says it shouldn't happen. It did. An agent with broad system permissions is a blast radius, and the blast radius doesn't shrink because the model is smart — if anything, a more capable model is more confident about the file operations it decides to run on its own. That risk doesn't disappear if you self-host, either. But when the model is yours, on hardware you control, you decide the permissions and the sandbox, and a mistake stays on your machine instead of somebody else's cloud account.
What this means if you run a small business
If you've been waiting for local AI to stop being a compromise, this is the week to look again. A quantized 20-to-30-billion-parameter model holding 90% of its original capability on a laptop is no longer a toy — it's realistically usable for drafting, intake, summarization, first-pass bookkeeping, anything where the data shouldn't leave the building. Healthcare intake, legal drafting, financial records: the privacy argument for running AI on-premise just got a lot cheaper to act on, because the hardware requirement dropped from "expensive workstation" to "device you already carry." You don't need a rack, and you don't need a monthly per-seat bill that scales with usage — you need one machine, configured once, that keeps doing the job whether it's used ten times a day or a thousand.
The one condition is non-negotiable, given what happened to GPT-5.6 this week: test any agent's write permissions in a sandbox before you let it touch a real file, local model or not. Full system access is a convenience feature that becomes a liability the first time the model is wrong about what it's doing, and "the model is usually right" is not a data-protection policy. Start read-only, watch what the agent tries to do for a week, and only widen the permissions you've actually seen it use correctly.
I've spent this year on my own version of this exact trade-off. A 123GB frontier model looked great on paper and ran out of memory every time I actually asked it a question on my 128GB Mac; a leaner 32-billion-parameter model just worked, cleanly, every single day since. Bonsai 27B is that same lesson pushed to the extreme — sometimes the smaller model that actually runs on the hardware in front of you beats the bigger one that technically exists on a spec sheet.
Hold me to this
So here's the claim, and it's checkable: by the end of August, at least one more open-weight release lands in the 20-to-30-billion-parameter range claiming 85% or more of frontier benchmark performance on hardware you can buy at a normal electronics store. If instead the next month of releases go back to being trillion-parameter-only, with no compressed counterpart shipping alongside them, then the two-track story I'm describing today collapses back into a single race — and I'll say so, right here, next week.