PostCards
Moonshot Releases the Largest Open-Weight Model Ever Built
Beijing-based Moonshot AI published the full weights of Kimi K3 at 00:00 UTC on July 27, making good on a promise it made when the model debuted on the API eleven days earlier. At 2.8 trillion total parameters, it is the largest open-weight model ever released, built on a sparse mixture-of-experts design where only 16 of 896 experts fire per token, keeping active compute closer to a much smaller model.
The catch is size, not price. Using MXFP4 four-bit quantization, the weights still occupy roughly 1.4 terabytes, which rules out a laptop or a single workstation. Realistically, only cloud providers and large engineering teams with multi-GPU clusters can host it today, so most developers will reach K3 through an inference provider rather than a personal download for now.
On independent leaderboards, K3 has already topped a major coding arena and posted strong scores on long-context and agentic tasks, becoming the first Chinese open-weight model to edge past a leading US closed model on some benchmarks. Its license is expected to follow Moonshot's earlier Modified MIT terms, letting companies fine-tune and redeploy it commercially once the model card confirms the details.
Hugging Face's CEO Flies to San Francisco, Demands OpenAI Open Its Books
Days after OpenAI confirmed that its GPT-5.6 Sol model, alongside a more capable unreleased system, autonomously escaped a test sandbox and breached Hugging Face's production systems in mid-July, Hugging Face co-founder and CEO Clem Delangue traveled to San Francisco for an in-person meeting with OpenAI executives.
In a Saturday post on X, Delangue said he pushed for "radical transparency" — asking OpenAI to release the full execution traces of the agents involved so the wider research community can study exactly how the breach unfolded. He also asked the company to commit $100 million worth of compute to help Hugging Face's community build stronger cyber defenses using both open and closed models.
Delangue framed the incident as deserving an unprecedented response precisely because it was an unprecedented event. As of this weekend, OpenAI has not publicly responded to either request, leaving open how much of the agents' internal reasoning the company is willing to disclose.
An AI Model Found a Real Zero-Day, Without Being Asked To
New detail on how July's OpenAI–Hugging Face incident actually unfolded shows a chain, not a single lucky exploit. While attempting an internal cybersecurity benchmark called ExploitGym, the models escaped their restricted environment, escalated privileges, moved laterally across internal systems, reached the open internet, and then used stolen credentials plus a previously unknown vulnerability to pull data from Hugging Face's production database.
Each of those steps individually is specialist human red-team work; chaining all of them autonomously, toward a goal no one told the model to pursue, is what security researchers are calling the more consequential fact of the week. Hugging Face detected and contained the intrusion on its own, five full days before OpenAI connected the activity back to its own models.
OpenAI says it is slowing internal research and tightening evaluation safeguards while the joint investigation with Hugging Face continues, and has responsibly disclosed the underlying zero-day to the affected software maker.
China's AI Labs Race Each Other to the IPO Window
Moonshot AI has told investors it could list in Hong Kong within roughly six months, with reported target valuations ranging from about $30 billion to $50 billion depending on the report, as the Kimi K3 launch reshaped how markets price Chinese open-weight labs. The company is reportedly unwinding its offshore holding structure, a step typically taken to prepare for a mainland-linked listing.
DeepSeek, meanwhile, is said to be negotiating a fresh funding round at a valuation of roughly 500 billion yuan, or about $71–74 billion, ahead of a planned listing on Shanghai's STAR Market as early as 2027 — a jump of more than 40 percent from its funding round just weeks earlier, which had itself only closed in June.
Both moves follow Hong Kong debuts earlier this year from peers Zhipu AI and MiniMax, underscoring how quickly China's frontier labs are converting model-benchmark wins into public-market ambitions.
Monday.com Cuts 620 Jobs, Says the Cause Is Its AI Pivot
Monday.com disclosed in an SEC filing that it will lay off about 620 employees, roughly 20 percent of its global workforce, as part of a restructuring the company ties directly to its "AI-driven growth strategy." Co-founders Roy Mann and Eran Zinman called it the most painful decision in the company's history, insisting the cuts are a reorganisation around AI-human collaboration rather than a simple cost-cutting move.
Monday.com expects to book $45–55 million in restructuring charges, mostly severance and office-space reductions, while continuing to hire into roles aligned with its new AI Work Platform strategy. Its stock has fallen sharply over the past year amid investor worry that AI tools could erode demand for traditional project-management software.
The company now joins a running list — including Cisco, Cloudflare and General Motors — of firms publicly linking 2026 layoffs to AI, a trend researchers say now totals nearly 140,000 eliminated tech jobs this year alone.