The Seventeen-Day Tide: Four Chinese Open-Weights Models, One Quiet Re-Drawing of the AI Map
Between April 7 and April 24, four Chinese laboratories released open-weights coding models that landed at roughly the Western frontier. The story is not the speed. The story is the geography.
In seventeen days, the centre of gravity of open-source artificial intelligence shifted east, and most of the world did not notice.
On April 7, Z.ai released GLM-5.1, a 754-billion-parameter mixture-of-experts model. On April 20, Moonshot released Kimi K2.6, a one-trillion-parameter vision-language system designed to coordinate hundreds of agents on a single task across multi-day plan-write-test-debug loops. On April 24, DeepSeek released V4 Pro, scaling its architecture to 1.6 trillion total parameters with 49 billion active. In the same window, MiniMax released M2.7, sitting strongly on the published intelligence-versus-size frontier.
Four laboratories. Seventeen days. All four with weights and inference code released under licences that allow downstream use without API gatekeeping. On the SWE-Bench Pro benchmark — a coding evaluation closer to real software-engineering work than older toy tests — the new Chinese stack converged at 58 to 59 per cent. Closed Western flagship models sit a few points higher; the gap is narrower than at any point in the last three years and is closing approximately every quarter.
This is not a story about who is winning. It is a story about which version of artificial intelligence is becoming generally available — and to whom.
What "open-weights" actually means for a citizen
The phrase is technical, the consequence is political. When a laboratory publishes the weights of a frontier model, it permits anyone with sufficient compute — and increasingly, that means anyone with a well-specified consumer GPU rig — to run the model locally, fine-tune it on private data, audit its behaviour, and deploy it without depending on a remote API or asking permission from the laboratory that built it.
The opposite stance — which has dominated the Western frontier for two years — is that frontier models should be served only through metered application interfaces, with the weights themselves treated as competitive trade secrets and, increasingly, as objects of safety regulation that ought not be released at all.
For users with capital, the difference is operational. For everyone else, the difference is political. An open-weights model is a model that can be deployed by a small business in São Paulo, a research lab in Lagos, a startup in Hyderabad, or a civic-tech collective in Manila without depending on a foreign company's price list or terms of service. It is the difference between a public utility and a private metered franchise.
The seventeen-day Chinese release window has, by sheer volume, made open-weights the default condition of frontier AI in two specific domains: agentic code generation and large-context multimodal reasoning. That is not a small change.
The four releases, plainly
GLM-5.1 (Z.ai, April 7). Mixture-of-experts at 754 billion parameters, of which a far smaller subset activates per token. The architecture is a continuation of the GLM family, which has been quietly improving release-by-release since 2023. The April 7 release was paired with a strong inference-cost story: at scale, the model is meaningfully cheaper to serve than comparable Western flagships.
Kimi K2.6 (Moonshot, April 20). A one-trillion-parameter vision-language model. The interesting design choice is the long-horizon agentic loop — Kimi K2.6 is built to coordinate plan-write-test-debug cycles that last for days, with the ability to instantiate hundreds of subordinate agents working on a single task. This is the design space that matters for autonomous engineering work, and it is where Moonshot has been concentrating its research.
DeepSeek V4 (DeepSeek, April 24). The headline number is 1.6 trillion total parameters with 49 billion active. The deeper story is that DeepSeek has, with each release, extracted more capability per dollar of inference than any other laboratory. V4 Pro is the most capable open-weights coding model that exists at the time of writing — narrowly, in specific evaluations, and at a fraction of the inference cost of the closed Western frontier.
MiniMax M2.7 (MiniMax, in the same window). A strong showing on the published intelligence-versus-size frontier — meaning it delivers more capability per parameter than its peers. MiniMax's positioning has historically emphasised cost-efficient deployment for application developers; M2.7 continues that emphasis.
Read together, these four releases are not duplicative. They cover different pieces of the same problem: scale (DeepSeek), efficiency-per-parameter (MiniMax), agentic coordination (Kimi), and architectural maturity (GLM). A practitioner choosing among them is choosing a stance, not picking a winner.
What this changes for the global AI economy
The pre-April situation, in plain terms, was that any economy that wanted to participate in the frontier of AI either had to pay Western frontier laboratories at API prices or had to accept second-tier capability for sovereign reasons. The post-April situation is that any economy with sufficient compute can run frontier-equivalent open-weights models locally. The cost structure is fundamentally different.
Three implications follow.
First, sovereign AI deployments become tractable for middle-income economies. A government that wants its agencies to run AI on data that cannot leave its jurisdiction now has a credible technical path. India, Indonesia, Brazil, Vietnam, the United Arab Emirates and several African states have all signalled in the last twelve months that this is a priority; the April releases give those signals technical substance.
Second, the developer-tooling economy diversifies. The closed Western frontier has, until now, been the default substrate for new AI-powered software products. The April release window has shifted the default among engineering teams from "build on a Western API" to "evaluate the open-weights stack first, fall back to closed if needed". This is not yet visible in headline market-share numbers, but it is visible in the procurement decisions of companies that are not VC-darling start-ups.
Third, the safety conversation changes shape. When frontier capability is concentrated in three or four closed Western laboratories, safety governance can be conducted as a small-N negotiation between those laboratories and their home governments. When frontier capability is also available in open weights from multiple jurisdictions, the safety conversation must become genuinely multilateral or it stops being effective. The April window forces that recognition.
The stance The Global Federation takes
We have written before that the open-source democratisation of AI is one of the most equitable technological developments in a generation, and we maintain that position. The seventeen-day Chinese tide is not a problem; it is, on net, an opportunity for every economy that has been priced out of the AI conversation by a closed-API regime denominated in dollars.
But three caveats deserve to be named.
The first is licensing nuance. "Open weights" is not a synonym for "open source" in the original software sense. Several of the Chinese releases come with use-case restrictions, deployment-region restrictions, or commercial-license requirements that more closely resemble the Llama family than the BSD-style licences that built the Linux kernel. These caveats matter, particularly for regulated industries and for any deployment that requires legal certainty about downstream redistribution. Read each licence carefully.
The second is geopolitical. Every laboratory operates within a regulatory environment. Models trained in one jurisdiction may carry training-data choices, content-moderation choices, and refusal behaviours that reflect that jurisdiction's preferences. These are, in a real sense, civic choices encoded into the model. A government, a company, or a civic-tech collective adopting an open-weights model from any jurisdiction — Chinese, American, or European — should be willing to invest in evaluation, alignment work, and behavioural auditing rather than treating the weights as a neutral utility.
The third is the safety frontier specifically. The capability gap between open-weights and closed-frontier has narrowed sharply on routine engineering tasks. It remains wider on the specific axes that frontier laboratories have invested heavily in: long-horizon agentic safety, refusal robustness against adversarial prompting, and provenance of training data. The right reading is that open-weights are now safe enough for ordinary deployments and are still maturing on the highest-risk ones.
What to watch in the next sixty days
Three signals will tell us whether the April window was a single tide or a regime change.
Whether Western frontier laboratories respond with their own weight releases. The Nemotron Coalition launched in early May, anchored by NVIDIA, has signalled an intention to release frontier-level open weights. If that coalition delivers in the next quarter, the open-weights frontier becomes genuinely multipolar rather than primarily Chinese. If it does not, the geographic concentration intensifies.
Whether sovereign AI procurement begins shifting visibly. Watch for announcements from Indonesia, Brazil, the UAE, Saudi Arabia, and India about which models their public sectors are running. The procurement decisions are the most reliable indicator of whether the open-weights tide is being absorbed into actual public infrastructure.
Whether developer tooling consolidates around an open-weights default. Tooling vendors — IDE integrations, agent frameworks, deployment platforms — typically lag behind model releases by three to six months. By August, the question of whether their default substrate has shifted from closed-API to open-weights will be visible in their documentation and their procurement contracts.
The seventeen days between April 7 and April 24 may turn out to be the single most consequential stretch for the global AI economy in 2026. The window did not produce a single dominant model. It produced four credible ones, all open, all from a single geography, all in less time than it takes most Western laboratories to ship a point release.
The map of who controls the frontier is being redrawn. The interesting question is whether the rest of the world will pick up what has been put on the table.
The Global Federation covers AI as a question of public infrastructure and global civic equity, not just a technology beat.