Two things happened this week that read like opposite bets and are the same bet paying off twice. PrismML's Bonsai 27B — a 1-bit and ternary build of Qwen3.6-27B — compresses under 4GB and holds 90% of its performance on an iPhone. The same week, Thinking Machines shipped Inkling at 975 billion parameters and Moonshot's Kimi K3 landed at 2.8 trillion.
The verdict: the frontier is not one race anymore. It is two, and only one of them needs a data center. A reasoning model in under 4GB on hardware you already own is not a demo. It is a deployment target.
Two expeditions, one base camp
Be precise about the difference, because it is not "sparse versus dense." Inkling activates 41 of its 975 billion parameters per token — the mixture-of-experts trick that keeps frontier inference affordable. Bonsai takes the same instinct further down: ternary and 1-bit weights to make 27 billion parameters fit where they have no business fitting. Soofi's 30B German-English model converged on the same efficient-by-design instinct in the same seven days. One expedition climbs; the other heads downhill toward you; both travel as light as the engineering allows.
What does brute force still buy? Real capability at the unsolved end: GPT-5.6 Sol Ultra reportedly proved a fifty-year-old conjecture in under an hour with 64 parallel subagents, and a Penn professor disproved a thirty-year-old statistics conjecture in ninety minutes. Impressive. Also not what a law firm drafting contracts or a clinic doing intake needs on a Tuesday. They are not short on mathematical horsepower; they are short on a model they can trust with the data.
The blast radius is the decider
The same week handed you the reason to care which side you are on: GPT-5.6 wiped users' home directories in Full Access Mode. OpenAI says it shouldn't happen. It did. An agent with broad permissions is a blast radius, and a smarter model is a more confident one about the file operations it runs on its own. Self-hosting does not remove that risk, but when the model is yours, you set the permissions and the sandbox, and a mistake stays on your machine instead of somebody else's cloud account. Whatever you run: start agents read-only, watch a week, widen only what you saw used correctly.
If you have been waiting for local AI to stop being a compromise, look again. A quantized 27B model at 90% capability on a laptop is realistically usable for drafting, intake, summarization, first-pass bookkeeping — anywhere the data shouldn't leave the building. Healthcare, legal, financial records: the on-premise privacy argument just got cheaper to act on, because the hardware requirement dropped from workstation to the device in your bag. One machine, configured once, no per-seat meter.
I ran my own version of this trade this year. A 123GB frontier model looked great on paper and ran out of memory on my 128GB Mac every time I asked it a real question. A 32B model just works, every day since. Bonsai is that lesson pushed to the extreme: the smaller model that runs beats the bigger one that technically exists on a spec sheet.
The claim: by end of August, another open release in the 20-to-30B range claims 85%+ of frontier benchmarks on hardware from a normal electronics store. If the next month goes back to trillion-parameter-only, the two-track story collapses and I'll say so here.