There's this thing happening in AI right now that nobody's really putting a name to. You can feel it in the product launches, in the way people are talking about inference costs, in the sudden appearance of tools that let you run models on hardware you already own. It's not quite a trend yet — more like a current pulling beneath the surface.
Two stories this week made me want to write about it.
Sigma Is What Happens When a Browser Takes Privacy Seriously
I spent some time with Sigma this week, and honestly, it's the first AI browser that didn't make me feel like I was being watched. Most of the "AI browsers" that launched over the past year — Perplexity's Comet, OpenAI's Atlas, Chrome with Gemini baked in — they all route your prompts through somebody's cloud. Sigma doesn't have to.
It ships with a local LLM called Eclipse that handles the chat side entirely on-device. No prompts leave your machine. But here's the pragmatic part — it's not purist about it. You can plug in cloud models when you need the heavy lifting, which means it's a hybrid, not a religious statement. That flexibility is what makes it interesting to me, not the privacy angle alone.
The browser itself is Chromium under the hood, so extensions work, sites render fine, and there's nothing to relearn. The difference is in the new tab page — you get three options: Search, AI Chat, and Agent. Search is just search. AI Chat drops you into a conversation with Eclipse. Agent is for handing off multi-step tasks that need actual browser actions. It's clean, it's fast, and the local model handles most day-to-day questions without breaking a sweat.
A lot of people are wondering whether local models can actually replace cloud APIs for everyday use. Sigma's approach suggests the answer is "not entirely, but more than you'd think." The local Eclipse model handles summarization, quick research, drafting — the kind of stuff you'd normally open ChatGPT for. The cloud option is there for when you need serious reasoning. It's a reasonable compromise.
One thing that bugs me: the passphrase setup on first launch is a genuine 12-word recovery phrase, and if you lose it, your Spaces data is gone. That's a real friction point for non-technical users. I get why it exists — end-to-end encryption doesn't work without it — but asking people to manage a BIP39 seed phrase before they can browse the web is… a choice.
The Model Apple Already Put in Your Mac
Sigma's local approach makes more sense when you realize Apple has been doing the same thing quietly for a year now.
Starting with macOS 26 Tahoe, every Apple Silicon Mac ships with the Apple Foundation Model — roughly 3 billion parameters, mixed 2-bit and 4-bit quantization, running entirely on-device across the Neural Engine and GPU. Apple just never gave you a way to talk to it directly. It's there for app developers to call through the FoundationModels framework, but there's no chat window, no settings toggle, nothing.
Then someone built Apfel.
It's a tiny MIT-licensed tool — one brew install apfel away — that wraps Apple's framework and exposes the model as an OpenAI-compatible server on port 11434 (same port Ollama uses, which is a nice touch). You get a CLI, an interactive chat mode, and most importantly, an HTTP endpoint that anything speaking the OpenAI protocol can use.
I tried hooking it up to Home Assistant, and it worked. My laptop, running a local model Apple already loaded onto it, answering smart home queries without a single cloud call. No API bill, no privacy concerns, just a MacBook acting as a free local AI server.
To be fair, it's a small model with a tiny context window. You're not going to run complex agentic workflows on it. Tool calling isn't really viable. But for summarization, quick Q&A, and simple automation? It's sitting there, idle, on millions of Macs. That's a weirdly large amount of free compute going unused.
Kimi K3's Pricing Is a Trap If You Don't Read the Fine Print
Moonshot AI launched Kimi K3 on July 16, and the numbers are impressive on paper. 2.8 trillion total parameters, 1 million token context window, native vision and tool use, and an API priced at $3 per million input tokens and $15 per million output tokens. Open weights promised by July 27.
But I did the math, and the headline doesn't tell the whole story.
First, that 2.8T figure is total capacity, not active parameters per token. K3 uses a Mixture-of-Experts architecture with 896 experts, activating 16 per token. Moonshot hasn't disclosed the active parameter count, which is the number that actually matters for inference cost and memory requirements. At 4-bit quantization, storing all 2.8T parameters would need about 1.4 TB of raw weight storage — before KV cache, runtime buffers, and replication. Moonshot recommends 64 or more accelerators for deployment. This is not running on your laptop.
Second, the weights aren't downloadable yet. Moonshot says they'll release them by July 27, but as of launch day, no checkpoint was available. That's a commitment, not a release.
Third — and this is the one that'll actually hit your wallet — early independent testing from Artificial Analysis shows K3 outputting roughly 130 million tokens during evaluation, versus a 63 million median for comparable models. More output tokens per task means that attractive $15/M rate doesn't go as far as you'd think. If the model is inherently verbose, your per-task cost could be double what the per-token price suggests.
I'd test K3 for long-context coding and research — the 1M context window is genuinely useful. But I wouldn't route production traffic through it without output caps and a task-level cost comparison first.
A Quick Note on the Security Side
TechRadar reported this week on something called "HalluSquatting" — an attack vector where bad actors exploit an LLM's tendency to hallucinate. By planting fake package names and documentation that LLMs are likely to recommend, attackers can create botnets out of AI-assisted coding tools. The researchers found that popular AI coding assistants can be tricked into suggesting malicious packages that don't exist but sound plausible.
It's a reminder that every new capability comes with a new attack surface. Running local models helps with privacy, but it doesn't automatically make you secure.
Look, I'm not saying local AI is ready to replace cloud APIs tomorrow. The Apple model is tiny. Sigma's Eclipse is capable but limited. Kimi K3 is still finding its footing on pricing and availability. But the direction is clear: more AI is going to run on devices we already own, and that changes the calculus around privacy, cost, and control in ways that the cloud-first narrative doesn't account for.
If you have an Apple Silicon Mac, install Apfel and see what your machine can do. Try Sigma if you're tired of every browser feeling like a data collection instrument. And keep an eye on the Kimi K3 weights drop on July 27 — if the active parameter count is reasonable and the license is usable, it could be a genuinely interesting open model.
Or maybe none of this matters and we'll all be back to using ChatGPT by August. That's the fun part about this space — nobody really knows.
If you found this useful, check out Decision Calculator — a tool I've been using to think through some of these trade-offs more clearly.
Top comments (0)