In partnership with
In Todayโs Issue:
๐ฆ The biggest open model ever built becomes a free download
๐ Anthropic breaks with the open-weights letter
โ๏ธ Washington has four days to define "frontier"
๐ก๏ธ Microsoft's cyber setup beats Mythos by 12 points
๐งฌ A chatbot out-ranks half the protein specialists
๐ค A robot lab puts more than 100 machines online for anyone to drive
โจ And more AI goodnessโฆ
โก The Signal
The most capable AI model anyone can now download for free is Chinese, and it sits close enough to the frontier to change the argument.
On Monday, Moonshot AI published the complete weights of Kimi K3, a 2.8-trillion-parameter model that scores 57 on the Artificial Analysis Intelligence Index, behind only Claude Fable 5 and GPT-5.6 Sol. None of it sits behind an API key. Anyone with the hardware can run it, inspect it, fine-tune it, or host it themselves. That lands in the same week Anthropic publicly broke with the industry's open-weights letter and Washington races toward an August 1 deadline to define which models are dangerous enough to require a government look before release. The technology keeps settling this argument faster than the policy can frame it.
All the best,
Kim Isenberg
(Anthropic)
๐ Anthropic Breaks Ranks on Open Weights
Anthropic published its own position on open-weights models on Monday, and it is not the one the rest of the industry signed. CEO Dario Amodei wrote that the company does not support banning open models as a category, but rejects the claim at the center of the industry letter now carrying roughly 50 names: that open models automatically make AI safer. Anthropic wants three things instead: tighter chip export limits on China, action against state-backed distillation, and safety testing tied to what a model can do rather than how it is licensed. "Open-weights models that don't have dangerous capabilities are a public good," the post argues.
๐ tl;dr: Test the capability, not the license.
(Getty Images via TechCrunch)
โ๏ธ Washington Has Four Days to Define "Frontier"
The White House is close to finalizing the voluntary framework that decides which AI models should be handed to the government for review before they ship. The Office of the National Cyber Director sent a draft to OpenAI, Anthropic and Google about two weeks ago, and the three returned a joint set of edits. The June 2 executive order set an August 1 deadline, and the White House has reportedly told some companies that reviews would be run by the NSA alongside CAISI, the small standards agency inside the Commerce Department. Two questions are still unresolved: how the framework defines a frontier model, and whether open-source models are treated differently from closed ones.
๐ tl;dr: The definition, not the technology, decides who ends up regulated.
(Microsoft AI)
๐ก๏ธ Microsoft Ships Its First Cyber Model
Microsoft launched MAI-Cyber-1-Flash on Monday, its first in-house security model, running inside a harness of more than 100 agents called MDASH. Paired with GPT-5.4, the setup scored 95.95% on CyberGym, a benchmark for finding and fixing real vulnerabilities in large codebases, roughly 12 points above Anthropic's Mythos 5. That is the model whose offensive cyber ability prompted June's executive order. Microsoft says the small model absorbs up to 90% of the workload, halving the cost of its own previous best configuration.
๐ tl;dr: The model judged too dangerous to ship just got out-scored by one Microsoft is selling.
๐ฌ Watch This
Sam Altman on building while the ground keeps moving.
In a 70-minute conversation with Ti Morse published Saturday, the OpenAI CEO spends less time on models than on operating inside chaos: trusting exponentials, keeping suppliers on OpenAI's timeline, texting 300 to 400 people a day, and shutting down Sora to concentrate on Codex. The striking moment is how flatly he says we are already inside the singularity, in the same register he uses to discuss buying compute. Worth watching for the operating philosophy rather than the forecasts.
"The biggest risk in AI is concentration of power."
โ Clรฉment Delangue, CEO, Hugging Face
Delangue's case is that openness levels the field rather than endangering it. Kimi K3's weights landing on his own platform this week is that argument made concrete.
Microsoft is reportedly so short of AI capacity that it has been evaluating renting compute from Amazon and Google, its two closest cloud rivals, with Amazon said to have stepped in after a run of GitHub outages. Internally, Copilot is served first and paying Azure customers get what is left.
(Cloud Computing News)
The Biggest Model Ever Built Is Now Free
The Takeaway
๐ Moonshot AI published Kimi K3's complete weights on Monday, the largest open-weight model ever released at 2.8 trillion parameters.
๐ It scores 57 on the Artificial Analysis Intelligence Index, third overall behind only Claude Fable 5 and GPT-5.6 Sol.
๐ A mixture-of-experts design activates about 104 billion parameters per token, with a 1-million-token context window and native text, image and video input.
๐ The real limits are hardware and paperwork: about 1.4TB of fast memory in four-bit precision, under a bespoke license with a clause for commercial hosts.
Until Monday, using a model this capable meant renting it. Moonshot AI has now handed the whole thing over, and the only gatekeeper left is whether you own enough memory to load it.
The company published the complete weights of Kimi K3, a 2.8-trillion-parameter mixture-of-experts model it first announced on July 16. Open weights means the trained model itself is public, so anyone can download it, run it on their own machines, examine how it behaves, fine-tune it for their own work, or serve it commercially. That is a categorically different thing from an API, where the provider keeps the model and can change, restrict, or withdraw it at any time.
(Artificial Analysis Intelligence Index, via Interconnects)
Kimi K3 scores 57 on the Artificial Analysis Intelligence Index, an aggregate of nine evaluations covering reasoning, coding and science. Only Claude Fable 5 at 60 and GPT-5.6 Sol at 59 rank above it, and both are closed. K3 sits above both Claude Opus 4.8 and GPT-5.5, and it tops the Frontend Code Arena leaderboard outright. The nearest open competitor on the index, GLM-5.2, sits six points back.
The architecture explains how a model this large stays usable. Of its 2.8 trillion parameters, only about 104 billion are active for any given token, so the cost of running it tracks the smaller number. It takes text, images and video natively, holds a 1-million-token context, and uses a new attention scheme Moonshot calls Kimi Delta Attention, which the company says makes long-context inference substantially cheaper than its previous approach.
(Kimi K3's attempt at Simon Willison's "pelican riding a bicycle" SVG test. Image: Simon Willison)
The frictions are real but narrow. In four-bit MXFP4 precision the weights still occupy roughly 1.4 terabytes of fast memory, which puts local hosting out of reach for individuals and into the range of well-equipped labs and clouds. The license is Moonshot's own rather than an off-the-shelf open-source one: it adds a separate-agreement requirement for companies running K3 as a paid hosted service above a revenue threshold, and an attribution rule for products above 100 million monthly users. For researchers, startups and anyone wanting to inspect a frontier-class model, neither restriction binds.
Why it matters: Access to frontier-class capability is no longer something a handful of companies control. Anthropic's position on open weights, Washington's August 1 definition of a frontier model, the question of who gets reviewed before release: every one of those arguments now has to contend with a 2.8-trillion-parameter model that is already downloaded and impossible to recall.
Sources:
๐ https://huggingface.co/moonshotai/Kimi-K3
๐ https://www.interconnects.ai/p/kimi-k3-the-open-weights-escalation
๐ https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3
How owning AI deployment expands your career
Across product, ops, and CX teams, a new kind of role is taking shape: the person responsible for making AI actually work, day to day. In this roundtable, three people living this shift share what it's really like: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. You'll hear how they carved out these roles, what the job looks like across industries, the skills they'd hire for, and the challenges they're tackling right now.
Watch the full conversation on demand.
The chart: PG-LLM asks a question nobody trained these models to answer: can a general-purpose chatbot predict how a single mutation changes a protein? Each model was scored by Spearman correlation across 217 ProteinGym assays, every one of them run at its highest supported reasoning setting. Claude Opus 5 leads every language model at 0.406, with GPT-5.6 Sol just behind at 0.402, and Kimi K3 is the best open model on the board at 0.314.
The lesson: These are chatbots, not biology tools, and they are already competitive with software built for this one job. Opus 5 clears the 0.36 median of the 46 sequence-only protein predictors and beats 49 of the 95 specialized comparators outright. General reasoning is transferring into hard science on its own, without a bespoke model for every domain.
The kicker: Kimi K3, the free download in today's lead story, is on this same chart. That means this capability is not sealed inside the closed frontier: anyone can pull the weights and start ranking protein variants tonight.
๐ค The Robot Lab That Wants You to Drive Its Robots
โก Bottom line: Israeli startup Enigma left stealth on Monday with $71 million and put more than 100 of its robots online for anyone to control.
๐ก Why it matters: Robot models have plenty of data on how machines move and almost none on how people actually give them instructions.
๐ What it means: The interface between person and machine is becoming robotics' hardest remaining problem.
Enigma's bet is that robots are already capable enough, and that the real obstacle is how awkward they are to instruct.
The company came out of stealth on Monday with a $71 million seed round led by Index Ventures and Ribbit Capital, with Sarah Guo's Conviction also taking part. Its founders, Jonathan Jacobi and Gal Niv, met in Unit 8200, Israel's signals-intelligence corps, and both arrived from cybersecurity rather than mechanical engineering. It shows in how they frame the problem.
(Enigma founders Jonathan Jacobi and Gal Niv. Photo: Natasha Zeriker via TechCrunch)
Their argument is that robot foundation models, the general-purpose models meant to let one machine handle many tasks, are being trained against the wrong bottleneck. Enormous effort goes into teaching robots to move. Almost nothing captures how an ordinary person actually tells a machine what they want. Jacobi's example is domestic: "If you had to do your dishes and spent 15 minutes explaining to a robot where to put everything, everyone reaches the point of 'Forget it, I'll just do it myself.'"
(Teleoperation rigs like this one are a common way to capture robot training data today. Photo: TechCrunch)
So Enigma built the data source it wanted. It runs more than 100 of its own robotic arms in hangars in Israel and California, and it has now opened them to the public, where anyone can take control in real time and have them paint, mix chemistry flasks, or fence with each other. Every session becomes training data about human intent.
It is a clever way to acquire that data and an unusually cheap one. The obvious objection is that people playing with robots over the internet for entertainment may not instruct them anything like someone would in a kitchen or a warehouse. Enigma is wagering that the gap is smaller than the cost of collecting the alternative.
Win AI Search Without a Big Team
92% of VCs use AI to find companies. 58% of buyers start there too. If you're not showing up in AI answers, you're invisible before the conversation even starts. Join HubSpot for Startups, Anthropic, and Marketing Against the Grain on July 30th (11am ET) for a live AEO teardown. Real startup. Real recs. Register and unlock the free Startup Visibility Bundle.