AI News Roundup — June 28, 2026

If yesterday had a single throughline, it was the quiet revenge of efficiency over scale — and of experience over hype. Tiny models punched far above their parameter count, Chinese labs kept eating into Western pricing power, and two sobering reality checks reminded everyone that an AI that can chat is not the same as an AI that can run a business — or build a car.

Small Models, Big Claims

The most practically useful news for anyone who runs models locally came from the lightweight end of the spectrum. Liquid AI shipped LFM2.5-230M, a 230-million-parameter open-weight model that hits 213 tokens/second on a Galaxy S25 Ultra and a still-usable 42 tokens/second on a Raspberry Pi 5. The standout detail isn't just the speed — it's the deployment surface. Day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX means this isn't a research curiosity; it's something you can drop into an existing edge stack without rewriting your inference layer. For sovereignty-minded builders, a model that runs on a Pi and beats larger competitors on instruction-following and data extraction is exactly the kind of cloud-independence primitive worth bookmarking.

The theme deepened with Sina Weibo's VibeThinker-3B, a 3B open model that reportedly rivals systems up to 333 times larger on math and coding benchmarks — achieved through multi-stage post-training rather than raw scale. The genuinely interesting finding is the asymmetry the researchers surface: logical reasoning compresses well into small models, but broad factual knowledge does not. That's a useful mental model for local deployment. It suggests the winning architecture for edge AI may be a compact reasoning engine paired with retrieval for facts, rather than a bloated generalist trying to memorize the world. Taken together, LFM2.5 and VibeThinker are evidence that the small-model frontier is where the most consequential open-weight progress is happening right now.

The Chinese Pricing Squeeze

The economic pressure on Western labs got more concrete. Coinbase has migrated to cheaper Chinese models like GLM 5.2 and Kimi 2.7, wiring up an automated router to pick the most cost-effective model per task. Combined with caching improvements that lifted hit rates from 5% to 60%, the company halved its AI spend even as token usage climbed. This is a meaningful signal: a major regulated US enterprise is now treating Chinese models as production-grade infrastructure, not a side experiment. For practitioners, the takeaway is architectural — the real leverage is in routing and caching, not loyalty to a single vendor.

The geopolitical framing came courtesy of Chinese cybersecurity firm 360, which launched AI security tools pitched directly against Anthropic's Mythos, with one already flagging 3,432 vulnerabilities in live deployments. Founder Zhou Hongyi candidly conceded Chinese models still trail Western ones by 20-30%, yet framed the contest as strategic "cyber-nuclear deterrence" — a national-priority push for independent capability. Read alongside the Coinbase story, the picture is clear: China is competing on price and self-sufficiency simultaneously, and that combination is what's putting the squeeze on incumbent pricing power.

Reality Checks: Where AI Still Falls Short

Three items formed an unintentional triptych on the limits of current systems. Princeton's CEO-Bench put AI agents in charge of fictional software companies for a 500-day simulation. Most went bankrupt; only three finished above starting capital — and a dumb rule-based heuristic beat nearly all of them. It's a brutal, clarifying result: today's models can produce plausible business prose but cannot yet sustain the compounding, long-horizon decision-making that running a company demands.

Researchers from Tencent and Chinese universities offered a diagnosis in their argument that AI won't become a true coworker until it stops answering and starts finishing tasks. Their prescription — persistent workspaces plus reusable skills — maps neatly onto why the CEO-Bench agents flailed: without durable state and accumulated competence, an agent restarts cold on every turn. The path from chatbot to colleague runs through memory and tool reuse, not bigger context windows.

Then Ford supplied the real-world version of the same lesson, rehiring veteran engineers after discovering that AI alone couldn't deliver quality products. Leadership openly admitted the error of treating automation as a substitute for seasoned judgment. The pendulum is swinging back toward AI-plus-human rather than AI-instead-of-human — a correction worth internalizing before you architect your own automation roadmap.

Infrastructure and the Money Trail

Follow the capital and you find Wall Street hunting the next Nvidia. The current favorite is Micron, positioned as an essential memory supplier riding explosive data-center demand. The thesis reflects investors broadening their bets across the entire AI supply chain rather than just the chip designers — a reminder that the compute economics underpinning every model you run start with memory and silicon, and that demand pressure there eventually shows up in your hosting bill.

Practitioner Toolbox

Two hands-on tutorials rounded out the day, both squarely aimed at people building rather than speculating. The first walks through building a stable Fable 5 Traces workflow in Colab, favoring manual JSONL parsing, tool-call normalization, secret redaction, and structure validation over fragile dependencies — a pragmatic recipe for producing clean chat datasets and trustworthy baseline training runs. The second is an OCRmyPDF pipeline in Python for turning image-only PDFs into searchable, archival PDF/A files, complete with Tesseract tuning, noise cleaning, orientation correction, and batch processing. Neither is glamorous, but both address the unglamorous reality of most AI work: getting messy data into a clean, auditable, locally controllable state before any model touches it.

The throughline: efficiency and rigor outperformed spectacle yesterday. Tiny open models proved reasoning fits in a backpack, Chinese alternatives kept reshaping the cost equation, and three independent reality checks reminded us that finishing tasks — and pairing AI with human expertise — still beats answering them alone.