The strategic change in 2026 is not that open weight models got cheaper. It is that several
Chinese laboratories now publish models close enough to the frontier that a UK or European business can
make a real choice between capability, price, control and operational independence. DeepSeek,
Alibaba’s Qwen team, Moonshot AI, Zhipu AI and MiniMax all pushed large open or openly deployable
families forward during the year. Moonshot’s Kimi K3 paper describes a 2.8 trillion parameter
mixture-of-experts model with 104 billion active parameters, native vision and a one-million-token
context window.
Nobody is being generous. The reasons are commercial and architectural, and both are worth
understanding before you build a procurement policy around them.
The commercial logic: commoditise your complement
If you sell the model, giving it away is cannibalisation. If you sell cloud, compute, chips or
enterprise integration, making the model layer abundant is an accelerant — cheaper models mean more
inference, and inference runs on the thing you actually sell. That structural asymmetry, not a difference
in values, explains most of who publishes weights and who does not.
Three other incentives line up alongside it. Published weights buy distribution through Hugging Face,
developer tooling, fine-tunes and third-party hosting in a way a less-trusted foreign API cannot. Models
that developers build around shape APIs, deployment stacks and ecosystem defaults. And constraints on
access to leading accelerators have sharpened the incentive to compete through efficiency — sparsity,
lower precision, attention efficiency, better serving software. The causal effect of export controls on
capability remains debated; the engineering focus on efficiency is plainly visible.
The architectural reason most coverage gets wrong
A dense model uses essentially all of its parameters for every token it generates. A mixture of experts
routes each token through only part of the network — a hospital with hundreds of specialists on staff, of
whom only a few are consulted for any one patient.
That produces two numbers that answer different questions. Total parameters mainly
determine how much memory is needed to hold the weights. Active parameters strongly
influence how much compute each token costs. Kimi K3’s 2.8 trillion total against 104 billion active is
why a model can be enormous on disk and still have an inference profile closer to a far smaller dense
model than the headline number suggests.
This is the mechanism behind the pricing, and it is why “how big is it?” is not a useful question on
its own. Our chapter on the open weight surge works through both numbers and what
each one actually costs you.
Handling the sovereignty argument honestly
There is a real continuity point here, and it is best made without geopolitics. On 12 June 2026 the US
government imposed export controls on two Anthropic frontier models, forcing a suspension of access while
user nationality could not be reliably verified. The controls were lifted on 30 June and service resumed
on 1 July.
Three weeks, and then resolved. Treat it as a precedent rather than an indictment: access to a closed
frontier service can change because of a decision outside your contract. That argues for a tested fallback
where the consequence justifies one — not for switching your stack on principle.
The distinction that saves the conversation
The single most useful thing you can do when this comes up in a policy discussion is separate two
decisions that sound identical and are not:
- Running open weights on infrastructure you control. The model is a downloaded
artefact. Your data does not leave your boundary. The risks are supply-chain risks: provenance, licence,
checksums, pinned versions, what the container pulls at runtime. - Calling a Chinese hosted API. Your data goes to a third-party processor in a
particular jurisdiction under particular terms. The risks are the ordinary ones you would assess for any
processor — residency, retention, subprocessors, legal access.
Policies that ban “Chinese AI” without making this distinction routinely block the low-risk option
while leaving the higher-risk one untouched, because the second is easier to buy on a card.
Where the low prices actually come from
Two forces, compounding. MoE sparsity, lower-precision arithmetic, efficient attention and better
serving stacks reduce the compute each token needs. Then third-party hosts compete to serve the same
downloadable weights, which pushes hosted prices down further.
Note the consequence for your own build: a self-hosted copy does not automatically beat those
economics, because your GPU can sit idle while a specialist provider pools demand across many customers.
Cheap open weights strengthen the case for using open models. They do not, by themselves,
strengthen the case for hosting them yourself.
One caution about writing any of this down
Do not turn this into a permanent leaderboard sentence in a strategy document. Releases, hosted prices
and licences move within weeks, and a dated table in a policy is worse than no table at all. Understand
the economics, then verify the live position before you procure. Our
guide to understanding AI in 2026 carries a dated market snapshot for exactly that
reason — and says plainly that it is a snapshot.