AI Foundationspredict · compress · act
act XI

The Open-Weight World

Open weights made the middle of the market cheap and left the frontier closed. That barbell shape is the real story — and it raises one question no licence can answer.

51

The Open Frontier

frontieras of 2026-09
before this →Running One Yourself

We can now see the whole shape of this act. Capable models are given away by labs in several countries, for four different reasons, under licences that range from generous to leashed, and they can run on hardware you own. What does that do to the field?

CLOSEDFRONTIEROPENCOMMODITYthe middle: good, not best — priced downbest capabilityyou pay, per callgood enoughyou host, for free
the barbell: the closed frontier and the open commodity both thrive, while the middle — good but not best — is squeezed.

The barbell

Commoditisation is what happens when a good becomes interchangeable: nobody can charge a premium for it. Every strong open model pushes another capability into the commodity end. The consequence is a barbell. At one end sit closed frontier models, where the best capability is sold per call. At the other sit open models that are good enough and free to host. The middle — a model that is good but not the best, sold as a service — is where the squeeze lands.

The split is measurable. On OpenRouter, a large public platform that routes requests to many models, closed systems from the big Western labs account for roughly 80% of all tokens processed; open models carry the rest. Look at spend rather than usage and the gap widens again — open weights win a substantial share of tokens while collecting a small fraction of the revenue. That is the barbell as a number, and it reads as a stable shape rather than a stage on the way to something else. Open weights won usage; the frontier kept the margin.

Reasoning in the open

The 2025 turn was reasoning. Once DeepSeek showed that reinforcement learning could produce it and released the weights, others followed: reasoning models appeared from Qwen, GLM, Kimi, and more, and small distilled reasoners put a thinking model on hardware that a year earlier ran a plain chatbot. Reasoning went, in months, from a closed frontier feature to an open one — a compression of the usual gap that is itself worth watching.

The question no licence can answer

open-weight risk

A guardrail added before release can be removed by anyone who has the weights. Fine-tuning can train safety behaviour away. A model cannot be un-released, and it cannot be patched for everyone at once. This is the sharpest open-weight risk: a public benefit and a public hazard in the same file. Set against it: open weights let more people study a model, and closed systems have their own failure modes — no outside scrutiny, decisions made privately. The honest position is that both ends of the barbell carry risk, of different kinds.

what to watchLicences drifting more permissive or more restrictive; evaluation scores that are really contamination; capability arriving at very small sizes; and whether any new kind of openness — data, not just weights — becomes normal.
RELEASEa capability leaves the lab
STUDYanyone can run and examine it
REMOVEanyone can also strip its guardrails

The closing thought

This act began with a question the first act posed in a different form: what is open? For weights, the answer is now concrete — a file, a licence, a machine that can run it. For the field, it is still contested: whether a model you can run but not rebuild is really open, whether the commodity end makes the frontier safer or the reverse, and whether a capability, once given away, can ever be recalled. Those are the open weights that no one has released yet.

introduces →commoditisationopen-weight riskbarbell
← previousRunning One Yourselfnext →The Harness