Open weights ship less often, and match the frontier anyway
Is AI actually getting more open over time?
Ask how open AI is getting and you will usually be shown a count of releases. By that measure the answer for 2026 looks grim: the first table below has open weights at their smallest share of any year on record.
That number is real, and it is close to meaningless as a scoreboard.
Release count measures how often a lab ships, not how good the models are. The proprietary side of 2026 is concentrated in a few labs that publish many increments, led by Google. Meanwhile Alibaba shipped as many open releases as proprietary ones in the same year, so even sorting labs into camps does not work.
The second table is the one that changes the picture. On context window — the only capability this dataset records on both sides of the licence line — the open side has not been behind since 2024. In 2025 the largest open-weights context window was Llama 4 Scout's ten million tokens, roughly ten times the largest proprietary one. In 2026 the two sit within a rounding error of each other.
So the honest reading is not that one side is winning. It is that the two measures disagree: open weights are a shrinking share of releases and are not a shrinking share of the frontier.
What would change my mind. Context window is one axis, and a cheap one to move — it says nothing about reasoning, coding or tool use, which is where the interesting argument actually is. This dataset records benchmark scores on thirteen records, far too few to say anything. Until that is fixed, treat everything above as a claim about context and licensing, not about intelligence.
How many releases, by licence
| Year | Open weights | Proprietary | Open share |
|---|---|---|---|
| 2022 | 4 | 1 | 80% |
| 2023 | 5 | 8 | 38% |
| 2024 | 14 | 18 | 44% |
| 2025 | 29 | 17 | 63% |
| 2026 | 61 | 123 | 33% |
| All | 113 | 167 | 40% |
The largest context window on each side
| Year | Largest open-weights context | Largest proprietary context |
|---|---|---|
| 2022 | 2,048 — OPT-175B ⚠︎ | 2,048 — PaLM |
| 2023 | 32,768 — Mixtral 8x7B | 100K — Claude 2 |
| 2024 | 256K — Jamba 1.5 | 2,097,152 — Gemini 1.5 Pro |
| 2025 | 10M — Llama 4 Scout | 1,048,576 — Gemini 2.5 Pro |
| 2026 | 1,048,576 — MiMo-V2.5 | 1,050,000 — GPT-5.4 |
What this does not claim
- Release count is not capability. The first table counts how often a model shipped, not how good it was. In 2026 the proprietary total is concentrated in a few labs that ship many increments.
- These are the 280 releases this dataset tracks, not every release that happened. Each year's direction is evidenced; the absolute counts are a sample.
access.open_weightsis recorded on every record and must agree with theopen-weightstag, so unlike most fields here this one has no gaps.- The share crossed the 50% line 3 times (2023, 2025, 2026). A single year's shift is not a trend.
- Context window is the only capability recorded on both sides here, and it is one axis, not a ranking. Open weights led it, or came within 5%, in 2022, 2025, 2026.
- Marked ⚠︎: OPT-175B (2022) — in the dataset, but the context window is not yet traced to a primary source. Treat any comparison resting on those as provisional.