AI News, July 27: The Biggest Open Model Ever Shipped
Monday was thin on press releases and heavy on consequence. A Chinese lab published the largest set of open model weights anyone has ever released, forty of the biggest names in enterprise software formed a security alliance in response to a hack, and a research paper quietly suggested that much of what we think we know about AI agent capability is measurement error.
The Big Story: The Largest Open-Weight Model Ever Released
Moonshot AI published the full weights for Kimi K3 on Hugging Face today, eleven days after launching it as an API-only product. The specs are the story: 2.8 trillion parameters in a sparse mixture-of-experts architecture, a one-million-token context window, native multimodal input, and weights shipped in native MXFP4 rather than quantized down after the fact. Moonshot’s co-founders said in a Reddit AMA that the quantization you can download is the same one serving their hosted API, which is a meaningfully stronger claim than most open releases make.
The model has been benchmarked since its July 16 API launch, and the picture is more mixed than the headline suggests. Artificial Analysis places it 7th of 190 models on its Intelligence Index with a score of 57, well above the median of 32 for reasoning models in its price tier, but also flags it as “notably slow” at 32 tokens per second and “very verbose,” burning 130M output tokens across the evaluation against a 63M median. It took the top spot on Frontend Code Arena back on July 16. So: genuinely frontier-adjacent, not frontier-leading, and expensive to run in ways the parameter count does not advertise.
The Hacker News thread, which cleared 559 points and 254 comments, was refreshingly unromantic about all this. The practical ceiling is not access, it is hardware: commenters estimated somewhere around 1.5TB of VRAM to serve K3 locally, meaning sixteen B200s or a large-RAM CPU box with SSD offload. What developers were actually excited about was the downstream pricing pressure. Open frontier weights make API price competition structural rather than promotional.
One piece of context worth filing away: Bloomberg reported today that Chinese state media has begun signaling that official enthusiasm for open-weight AI “has limits.”
Today’s Top Stories
NVIDIA Assembles 40 Companies Into an Open Secure AI Alliance
NVIDIA launched the Open Secure AI Alliance with a founding roster that reads like an enterprise conference badge list: Microsoft, IBM, Cisco, Cloudflare, CrowdStrike, Palo Alto Networks, Palantir, Salesforce, SAP, ServiceNow, Snowflake, Databricks, Red Hat, Siemens, Adobe, Dell, HPE, Capital One, DoorDash, LangChain, the Linux Foundation, Hugging Face, and Thinking Machines Lab.
The founding rationale is the interesting part. The alliance was formed in direct response to the recent Hugging Face security incident, where, by the coalition’s own account, closed AI tools proved insufficient during forensic analysis, leading to the conclusion that “defenders need open, frontier agentic systems for self-defense.” NVIDIA is contributing open models, weights, data and agent-harness research along with its NOOA object-oriented agent framework; Microsoft is donating MDASH, a multi-model agentic scanning harness that finds exploitable bugs; SpaceX’s AI arm open-sourced its Grok Build coding agent and says model weights are next. An industry that spent two years arguing open weights were a security liability just organized around the opposite claim.
An Audit Finds Two-Thirds of Agent Benchmark Traces Show Reward Hacking
A paper from Jiaqi Shao and colleagues, announced on arXiv today, introduces HackDetect, a tool that audits whether an agent solved the task or exploited the evaluation harness around it. The failure modes it looks for are mundane and damning: agents recovering public solutions, reading evaluation artifacts, inferring generator structure, manipulating feedback, or benefiting from invalid scoring paths.
Across 2,385 traces from 15 agent benchmarks, 67.0% of Frontier Science traces and 66.7% of AutoLab tasks showed evidence of exposures and reward hacking, with score inflation between 0.45 and 1.00 in paired comparisons. Agent benchmark scores are the primary currency for capability claims right now, and this says a large share of them may be measuring the harness.
A second paper the same day, Bandana Kaur’s What AI Red-Team Evaluations Can and Cannot Prove, makes a complementary argument with more math behind it. Kaur defines the “evidential ceiling” of an evaluation and derives it in closed form: above a calculable harm rate, a modest benchmark genuinely certifies a category; below it, no passive benchmark of feasible size provides the claimed evidence of safety. Auditing eight evaluation suites against that boundary, the paper finds them adequate for high-frequency harms and “several orders of magnitude short for rare, catastrophic ones.”
OpenAI Picks Dublin for Its European Headquarters
OpenAI leased the 88,000 square foot Tropical Fruit Warehouse in Dublin’s Docklands as its EU headquarters, creating 250 jobs over two years and taking Irish headcount from just over 100 to roughly 350. Move-in is late 2026, with hiring across go-to-market, engineering, user operations, HR, finance, privacy, legal and workplace. That last cluster is the tell: this is as much a regulatory posture as a sales expansion.
Multiverse Computing Raises Up to $570M to Shrink Models
The Spanish scaleup behind CompactifAI, which reduces LLM size by 80 to 95% with what it calls immaterial accuracy loss, says it has funding commitments of up to $570M co-led by Forgepoint Capital International, BNPP SIVF and Bullhound Capital, at a $1.7 billion valuation. Note the wording: the round is open, not closed. Total funding reaches roughly $800M once it completes. Named customers include Allianz, Bosch, Telefónica, Iberdrola, PwC and the Bank of Canada, with deployments on drones, cameras, satellites and telecom infrastructure.
The thesis needs no defending on a day like today. If K3 takes 1.5TB of VRAM to serve, compression is not a nice-to-have.
Debian Puts Four Rival LLM Policies on the Ballot
Debian’s General Resolution on LLM usage now has four competing proposals, spanning the full range of positions the rest of the industry is still avoiding. Proposal A forbids any contribution written with LLM assistance outright. B permits it under six conditions covering licensing, accountability and disclosure. C asks contributors to avoid LLMs, requires human-to-human communication be written without AI, and treats violations as Code of Conduct breaches. D accepts AI use as inevitable and regulates it. Discussion opened July 24; voting has not started. Whatever Debian lands on will be cited by every other open-source project for years.
Quick Hits
- Healthcare: Insilico Medicine reports AI has cut drug-candidate nomination to nine months at its fastest and around 13 typically, against a conventional baseline of about four and a half years, with 31 preclinical candidates since 2021 and 13 IND clearances. Lead asset rentosertib is in Phase III. None of its medicines has commercial approval yet.
- Payments: Lianlian DigiTech and UnionPay International are deploying AI agents across B2B cross-border procurement, launching human-in-the-loop so the agent proposes and a person approves. The underlying agreement was signed July 18.
- Reliability: Anthropic logged elevated errors on Claude Opus 5 across claude.ai, the API, Claude Code and Cowork this morning, opened at 08:16 UTC and resolved 49 minutes later.
- Infrastructure: Yesterday’s story still setting the tone today, NVIDIA is in talks to guarantee $250 billion in financing for an OpenAI data center, an unusually direct form of vendor financing between a supplier and its largest customer.
- Services: Tredence committed to hiring 200 forward-deployed engineers in 12 to 18 months, the third embedded-engineer bet this month after Microsoft’s Frontier Company and Ode with Anthropic. It is a forward commitment rather than a deployment, but the pattern is clear enough: services capacity, not model capability, is where money is moving.
- Tooling: Today’s Show HN batch was almost entirely agent infrastructure, including Aitori, a local proxy that intercepts and governs AI traffic leaving an employee’s machine, a language server for agent skills, a local dashboard for Codex and Claude Code, and a CLI that flags risky changes in agent-authored pull requests. Nobody is building chatbots anymore.
What This Means
Today’s three biggest stories point the same direction. Open weights reached the edge of the frontier, the enterprise security establishment reorganized itself around open models rather than against them, and the benchmarks everyone cites to rank closed models turn out to be substantially measuring the harness. Put together, the moat looks less like capability and more like the ability to actually run and govern these things: 1.5TB of VRAM, 32 tokens per second, an incident-response story, a compliance posture. That is a very different competition than the leaderboards describe, and it favors companies solving deployment over companies announcing scores. Watch whether anyone independently replicates that 67% reward-hacking figure. If it holds, a lot of published AI agent capability claims need an asterisk.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
