The largest open model is one almost nobody can run

Open weights were supposed to distribute capability. At 2.8 trillion parameters the licence is still open and the practical access is not — which turns openness from a question about permission into a question about who owns enough silicon to exercise it.

Moonshot completed the open-weight release of Kimi K3 at 2.8 trillion parameters, roughly 75% larger than DeepSeek V4-Pro. Anyone may download it. Very few can serve it, and that gap is the whole point.

Openness has two meanings and they have come apart

For most of the open-weight era they coincided: a model you could download was a model you could run, on hardware a university or a serious hobbyist could reach. At frontier scale that stops being true. The licence still grants permission; the infrastructure decides who can act on it. Open weights at 2.8T distribute the right to participate to everyone and the ability to a dozen organisations.

The release cadence compounds it. Seven models shipped between 17 and 23 July, three Qwen releases inside a single 72-hour window. Independent evaluation takes weeks. When releases outpace assessment, the analysis that reaches buyers is increasingly the vendor's own — which is a strange outcome for a movement built on verifiability.

Where the value actually lands

Not with the labs releasing the weights, and not with the enterprises downloading them. It lands with whoever can serve a 2.8-trillion-parameter model economically. That is a hosting business, and the capital markets have already worked this out — two $1.5B rounds into inference serving inside a single quarter is that thesis being funded.

The part worth arguing about

There is a reading where this is fine: weights being public means they can be audited, forked and preserved even if only large operators serve them. There is another where an open model nobody can run is open in the way a library with no doors is public. Both are defensible, and the policy debate has not caught up to the fact that the question changed while it was being had.

The frontier of open weights is now set at a scale that excludes almost everyone it was meant to include.

Taskade — Open-Source LLM History: GPT-2 to Kimi K3 (2026) → · Digital Applied — Seven Days, Seven Model Releases: The New AI Normal →