Open Weight AI Models: Open vs Closed by the Record

The Thesis

The story everyone tells about open-weight models is a race: China’s labs closing on America’s, open catching up to closed, the gap narrowing every quarter until it disappears. The coverage record does not support it. It supports something stranger and more useful.

Across 187,000 articles, model-release coverage splits into two camps — releases where the weights are published and anyone can download them, and releases that reach users only through an API or a product. Since 2024 the open camp’s share of that coverage has been 37%, then 36%, then 33%. Three years, three points of drift, through DeepSeek’s arrival, through Llama’s wobble, through the entire Chinese open-model wave. About one release story in three is an open-weight release, and it has been about one in three the whole time.

That stability is the finding, and it survives the thing that usually breaks such numbers: volume. Total release coverage more than doubled over the same span. Both camps grew hard. Neither took share.

What did change is the register. In 2026 open weights stopped being a technical preference and became a thing governments are asked to rule on — and the labs that publish nothing started lobbying about the labs that publish everything. A ratio that would not move on capability is now being contested on policy. That is the trade: not who wins the benchmark, but who gets to write the rule while the benchmark stays tied.

The Evidence

Start with the count that refuses to move.

The underlying figures: in 2024 the archive holds 94 open-weight release stories against 159 closed. In 2025, 116 against 203. In 2026 to date, 122 against 251. The absolute numbers climb steeply on both sides and the ratio sits still. The one year that looks different is 2022, when open led 12 to 7 — a majority built on nineteen articles total, before the release cycle existed at the scale it does now, and not a baseline anything should be measured against.

Within 2026 the quarterly detail says the same thing from closer up. Open-weight releases ran 34, then 41, then 47 across the first three quarters — a steady climb. Closed ran 74, then 107, then 70. The closed camp is the volatile one, spiking in the second quarter and falling back; the open camp just grinds upward. Nobody is pulling away.

Where the record does show a break is in what the open camp is releasing. Moonshot released Kimi K3, a 2.8-trillion-parameter model it said rivals Claude, and then published its weights under a bespoke licence. Alibaba said its 2.4-trillion-parameter Qwen3.8-Max tops Kimi K3 on some benchmarks. Thinking Machines Lab debuted Inkling, an open-weight mixture-of-experts model with 975 billion total parameters. Three years ago the open camp’s argument was that a small model you control beats a large one you rent. It has stopped making that argument, because it stopped needing to.

Capability claims are cheap; the security evidence is not. Researchers found that GLM-5.2 matches US models at finding security bugs — an open-weight model performing at frontier level on a task with a hard, checkable answer. A separate analysis cut the other way, finding that open models lag on cyber capability. Both are in the record, both are real, and the disagreement is the honest state of it.

The most-covered names in the open camp are not the ones the China narrative would predict. Meta leads at 131 articles, then Google at 69, Nvidia at 47, Mistral at 33, Microsoft at 31. The explicitly Chinese cluster — the entity “Chinese AI” at 27, DeepSeek at 17 — is real and rising, but the open-weight record is still substantially an American corporate record. The story of open weights is a story about what large US companies chose to give away.

The Companies

Meta is the largest single publisher in the open camp and the clearest case of a company changing its mind in public. In June 2025 the record has executives discussing “de-investing” in Llama; by July that year Zuckerberg was talking about needing to be rigorous about mitigating the risks of what Meta released; by December, reporting suggested a new model codenamed Avocado might not follow the old pattern. Then in August 2026 Meta released Muse Glimmer as an open-weight model with more planned. A retreat, and a return. Anyone treating Meta’s openness as a fixed property of the company is reading one frame of a film.

Google runs the opposite pattern: two products, two doors, no drama. Gemini goes out closed; Gemma goes out open. It launched Gemma 4 in April as its “most intelligent” open family and followed with DiffusionGemma, a 26-billion-parameter experiment, in June. Google is the only company in the record consistently shipping serious models through both doors at once, which makes it the least exposed name here to whichever way the rule goes.

Nvidia sits third in the open camp on article count while selling to everyone in both, and its position is the one that reads most clearly through Hamilton Helmer’s counter-positioning: it debuted Nemotron 3 Super, a 120-billion-parameter open-weight model, not because it needs a model business but because every open model that runs well on its hardware widens the market for the hardware. Open weights commoditise the layer directly above Nvidia’s. That is a structural reason to publish, and it does not depend on believing anything about openness.

The names absent from this list matter as much. OpenAI appears in the open-weight record 21 times, mostly for GPT-OSS and for saying it might do more. Anthropic appears 17 times, mostly for arguing about policy rather than shipping weights. Neither is public, and the closest public proxy for that camp — the cloud platforms that resell their models — is tracked under the inference layer rather than here.

The Lenses

Helmer’s 7 Powers asks what a company holds that a competitor cannot copy at acceptable cost. Publishing weights is the deliberate destruction of exactly that, in one specific layer, by someone who profits from the destruction. It is counter-positioning aimed downward: Meta and Nvidia give away the thing OpenAI and Anthropic sell, because neither of them sells it. The response of an incumbent facing counter-positioning is famously not to match the move — matching it destroys the business being defended — but to argue that the move should not be allowed. Which is what the 2026 record shows happening.

Shiller’s narrative economics explains why the ratio can hold flat while the noise rises. A stable one-third share is not a story; “China is winning” and “open models are dangerous” both are. Narratives spread on contagion, not on accuracy, and the coverage record is where you can see the two come apart — the share sitting still while the volume of argument about it climbs. Any position sized off the narrative rather than the ratio is sized off the wrong number.

What Moved

Z.ai launches GLM-5, calling its flagship open-weight model best-in-class — the first 2026 release where a Chinese open model's headline claim is frontier parity rather than value for money.

Google launches Gemma 4, its "most intelligent" open family, keeping both doors open while Gemini stays closed.

Researchers say GLM-5.2 matches the latest US models at finding security bugs — an open model at frontier level on a task with a checkable answer.

Moonshot releases Kimi K3 at 2.8 trillion parameters, saying it rivals Claude; the weights follow eleven days later under a bespoke licence.

Commentary in the record argues Chinese open models are proper competition to be welcomed rather than suppressed — the first sign the argument has become the story.

Technology companies sign a letter defending open-weight AI and urging lawmakers to avoid restrictions.

Reporting says OpenAI and Anthropic are quietly lobbying Washington regulators to restrict open models.

Dario Amodei says Anthropic has "never" backed an open-weights ban — a denial that only exists because the accusation did.

Reporting says the US AI framework excludes open models and defines a covered frontier model in a way that leaves them outside it.

Meta releases Muse Glimmer as an open-weight model with more planned, closing a fourteen-month wobble that began with talk of de-investing in Llama.

Qwen3.8-27B demonstrates an open-weight general-purpose model holding long context — capability arriving at the small end, where deployment economics bite.

The cluster is unmistakable. Ten of the eleven entries above fall in 2026, and seven fall in a five-week window from mid-July to mid-August. Within that window the letter, the lobbying report, the denial and the framework carve-out all land — four policy events in twelve days, against a release ratio that had not moved in three years.

What to watch is whether the ratio finally breaks, and in which direction, now that it is being argued about rather than competed over. The record’s own answer so far is that argument has not moved it at all.

Sources

Drawn from the TEXXR archive: 675 articles tagged as open-weight coverage and 1,313 as closed-weight, each classified by similarity to a defined centroid and then adjudicated individually. Release counts use the archive’s event classification to separate launches from business and policy coverage. Article IDs resolve at texxr.com/<id>.