Moonshot AI has released the full weights of Kimi K3.

The model has 2.8 trillion parameters, with 104 billion active at a time. It is far too large for an ordinary laptop. Its size may even limit adoption, according to researchers quoted by Nature.

The release makes the model portable.

A cloud provider with enough hardware can host K3. A research team can inspect it. A company can adapt it without sending every request back to Moonshot, subject to Kimi's licence and local law.

Closed AI companies sell a different arrangement.

They keep the model on their servers. They decide the price, supported countries, safety rules, available versions and whether your request reaches the model at all.

Open weights weaken that control.

That is why the fight around open AI is also a fight over business.

I am Alex, welcome to ShortCu8 by Innov8.

Lets Dive Deep 🐰

When someone calls an open model dangerous, separate three questions:

Mixing these together makes every open model look stolen or unsafe even though each problem needs a different rule.

Training a frontier model costs a fortune. Closed labs recover that money by keeping the model behind an API or subscription.

That gives them control over:

Training and hosting remain expensive.

The original company stops being the only gate.

Microsoft, OpenAI, Google, Meta, NVIDIA and dozens of other companies signed a July 24 statement supporting open-weight AI. Their argument is practical: open weights reduce dependence on one provider, increase competition and let organisations run models on their own infrastructure. The published list omits Anthropic and gives no reason.

Anthropic says Chinese labs used Claude to improve their own models through industrial-scale distillation.

Distillation itself is normal. A stronger model produces examples that help train a smaller model. Anthropic uses the technique too.

The allegation concerns how the access happened.

Anthropic says DeepSeek, Moonshot and MiniMax generated more than 16 million Claude exchanges through roughly 24,000 fraudulent accounts. It attributes more than 3.4 million exchanges to Moonshot and says the campaign targeted coding, reasoning, computer use and vision.

Those are Anthropic's allegations. They have not been independently settled in public.

If a competitor used fake accounts, proxies and regional-access evasion to extract proprietary capabilities, that conduct deserves investigation.

The alleged misconduct concerns how Claude was accessed. Publishing the resulting model files is a separate act.

Anthropic's safety concern is real.

Once powerful weights are released, the original company cannot reliably recall every copy, enforce one safety policy or track every modified version.

The commercial benefit of the same restriction is also real.

If weights stay private, the company keeps control over access, pricing and distribution.

The useful policy line is therefore narrow:

The current US position appears to be moving toward that distinction. Axios reports that the White House is backing open weights while threatening action against covert, industrial-scale extraction through fraudulent accounts or evasive access.

That is more precise than banning the model format.

Kimi K3 requires data-centre hardware.

Open weights can still matter here.

An Indian provider could potentially host the model inside India, subject to the licence, infrastructure cost and Indian law. Builders could then access it without waiting for the original company to support India or preserve access indefinitely.

The problem changes from:

to:

India can work directly on compute, hosting, audits and local rules.

Kimi K3's official model card publishes the full weights under a custom licence and describes a 2.8-trillion-parameter model with a one-million-token context window.

Anthropic openly describes distillation as legitimate while alleging that Moonshot and others crossed the line through fraudulent accounts and access evasion.

The new US industry statement makes the same distinction: legitimate distillation should remain available, while unlawful extraction should face targeted legal action.

The argument now concerns who controls intelligence after it has been built.

Now go and try opus 5 its kind of nice 🙎🙎

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Sign up for a free trial of Attio, the agentic CRM.

Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.

Review your pipeline in Claude, synced live from Attio via MCP.

That's it.

Try Attio Now

#AD2

Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.

Try Wispr Flow free