.svg)
parameters
parameters activated by inference
tokens of context window
On July 16, MoonshotAI, a Chinese research lab, released its new model Kimi K3, after Kimi 2.6 and Kimi 2.7 — both already frontier-adjacent models. K3 is the first open-weight model to land inside the frontier cluster: third on the Artificial Analysis Intelligence Index, behind only GPT-5.6 Sol Max and Claude Fable 5, and first in the Frontend Code Arena [1]. It is built for agentic and long-running tasks, as expected from a top contender. But the decisive difference is distribution: the weights are entirely free and published on Hugging Face, and at 2.8T parameters K3 overtakes DeepSeek V4 as the largest open-weight model [2]. Weights and technical report landed on 27 July, putting a Chinese lab on top of the open-weight field [3].
So what does it take to run frontier-level intelligence locally? The model has 2.8T parameters but activates only 104B on each inference run, thanks to its mixture-of-experts architecture (16 experts out of 896), and it supports a 1M-token context window. All of this is built on the lab's new attention architecture, Kimi Delta Attention [3, 4]. It currently ships under a modified MIT licence, without the original training data or code: open-weight, not open-source. The download alone is close to 1.4 TB, and self-hosting it demands colossal cluster infrastructure, roughly 2 TB of VRAM [5]. The most capable open-weight model in existence is therefore downloadable by anyone and operable by almost no one. The practical floor is not an enthusiast with a workstation; it is a funded organisation with an AI-oriented data center.
On 24 July, Jensen Huang published his first X post to circulate a three-page letter, "Open Weights and American AI Leadership" [6]; fifty companies, among them Mistral, Ollama, IBM and AMD, signed it within two days [7]. None of the arguments in it are new:
What is new is the audience. The case for openness stopped being a preference among practitioners and became a filing addressed to lawmakers — four days after reports that Washington was reviving a push to ban Chinese models, and eight days after the API debut of the leading open-weight model, which had just reached frontier level [8]. The letter was written during that API debut, but it landed three days before the weights went public. The original twenty-five signatories on day one included no company selling access to a closed frontier model. OpenAI and Google signed only after the omission was noticed; Amazon has still not signed, and Anthropic answered separately, in a related post by its CEO, Dario Amodei. Everyone here is talking their book: NVIDIA sells the accelerators, and the frontier labs need to stay rare. Neither fact makes either case wrong. With Kimi K3 benchmarking near the proprietary systems, the question is no longer whether an open model can do the job; it is whether you will still be allowed to download it.
Anthropic has now published its answer, and it is narrower than the framing it received. On 27 July, Amodei [9] states plainly that Anthropic has never advocated a ban on open-weight models, and calls open models without dangerous capabilities a public good. He agrees with much of the letter's thesis on access, competition and customer control. What he rejects is one claim in particular: that open weights necessarily make AI safer, or hand defenders an advantage over attackers [10].
He points instead to other choke points: keeping advanced chips out of authoritarian hands, stopping industrial-scale distillation (training on the outputs of a competitor's model), and mandatory safety testing for every sufficiently capable model, open or closed [11]. The last is the one worth sitting with, because it does not sort models by licence at all: where the letter treats every model release the same, Amodei sorts models by capability.
