Open weights took nearly half the top 12 in August. Will September keep up?

By the end of August 2026, a snapshot of Code Arena’s top 12 looked less like a closed-model club. Five entries carried non-proprietary license labels: Tencent’s hy4-preview at No. 5; Alibaba’s qwen3.8-flash-next and qwen3.8-27b at Nos. 8 and 12; and Z.ai’s glm-5.3-max and glm-5.3-flash at Nos. 10 and 11. That is nearly half the table—not half of every LLM ranking, and not five models with identical rights to use or redistribute their weights. Still, it is a striking amount of competition beside the closed systems near the top.

The prices make the snapshot harder to shrug off. Code Arena lists hy4-preview at $0.83 per million input tokens and $2.50 per million output tokens; glm-5.3-flash at $0.15 and $0.50; and qwen3.8-flash-next at $0.16 and $0.47. Several placements are preliminary, and leaderboard scores do not tell you how a model will perform on your own work. But these are no longer merely bargain-bin alternatives several generations behind. They are appearing within dozens of points of the leaders on this particular board.

The vocabulary matters here. Open-weight is not a synonym for open source, and a license label alone cannot settle every question about hosting, modification or commercial deployment. Developers need to read the terms and check that usable weights are actually available. When they are, the practical difference from an API-only model is substantial: a team can run its own inference, fine-tune or quantize the model, and build without routing every request through one provider. That matters most when data is sensitive, usage is steady and switching costs have years to accumulate.

Why release the weights at all? Because the API bill is only one way to win. A capable model that developers can adapt attracts integrations, optimizations and applications its creator would never build alone. If it becomes the default starting point for products and private deployments, giving up some direct token revenue may be a good trade. That is a strategic explanation for the releases, not proof that every lab is following the same playbook.

Nor does an open-weight surge require the most expensive frontier systems to open up. Labs can keep their highest-end models behind APIs while releasing smaller, faster or less cutting-edge members of the family. The result is not a clean victory for one ideology over another. It is a business model with two lanes: premium access to a closed frontier, and broad distribution through weights that others can run.

So what should we expect in September? There is no reliable industry vote to count here, and a crowded August does not make another crowded month inevitable. The cautious bet is that more useful weights will arrive, but that the labs with the strongest closed flagships will not rush to publish them. Watch for models that are good enough for a wide range of work—especially coding and agent tasks—and for licenses that make deployment feasible rather than merely exciting on launch day.

The case for another busy month is straightforward: price pressure, active Chinese labs, demand for deployments under customers’ own control, and the network effects of a developer ecosystem. The brakes are real, too: training costs, misuse concerns, export rules and the value of keeping the best reasoning capabilities exclusive. September could bring fewer releases than August and still tell us more. The revealing questions are whether the licenses hold up, inference is affordable, benchmark results travel to real tasks, and the models work well outside their creators’ clouds.

September could bring fewer releases than August and still tell us more.

For builders, the immediate lesson is to avoid designing an LLM stack around the assumption that two or three closed APIs will always be the only serious options. Keep prompts portable, make room to route between models, and evaluate candidates on your own tasks. Count the cost of dependence on a provider alongside the cost per token. August made open weights impossible to dismiss as a hobbyist sideline. September will test whether they are becoming a durable second route to capable AI—or a carefully managed complement to the closed frontier.