# Data notes

The snapshot and timeline start dates live in the canonical dataset, `data/releases.json`; every plotted event embeds primary-source links and an access qualification. The selection covers material general language/reasoning launches, not every model checkpoint or service update.

## Primary source indexes

| Provider | Source anchors |
| --- | --- |
| OpenAI | [Model release notes](https://help.openai.com/en/articles/9624314-model-release-notes), individual OpenAI launch posts |
| Anthropic | [Newsroom](https://www.anthropic.com/news), [API release notes](https://docs.anthropic.com/en/release-notes/api) |
| xAI | [News](https://x.ai/news), [company timeline](https://x.ai/company) |
| Google | [Gemini model posts](https://blog.google/innovation-and-ai/models-and-research/gemini-models/), Gemini API release notes |
| Meta | [AI blog](https://ai.meta.com/blog/), official Meta model cards |
| NVIDIA | [Nemotron research](https://research.nvidia.com/labs/nemotron/projects/), NVIDIA launch posts and model cards |
| DeepSeek | [API release history](https://api-docs.deepseek.com/updates), official DeepSeek repositories |
| Z.ai | [Model launch blog](https://z.ai/blog), official THUDM/Z.ai repositories |
| MiniMax | [News](https://www.minimax.io/news), [research](https://www.minimax.io/research), MiniMax-authored launch statements |
| Huawei | [MaaS release log](https://support.huaweicloud.com/wtsnew-maas/index.html), Huawei Cloud announcements |
| Tencent | [Hunyuan model log](https://cloud.tencent.com/document/product/1729/97765), [TokenHub log](https://cloud.tencent.com/document/product/1823/130675), corporate launch posts |

## Date and inclusion decisions

- Public access means an openly obtainable app, paid API, or model weights. An invitation-only beta is contextual. Public preview access counts.
- Dates use the earliest documented access, not necessarily the corporate publication date. Claude 3.5 Sonnet uses its June 20, 2024 API release; the launch post currently carries June 21. GPT-6.1 Sol uses the September 29, 2026 DevDay recap documenting public availability.
- Google Gemini 1.5 Pro uses April 9, 2024 public preview access, rather than the earlier private preview announcement. Claude 3.5 Haiku uses November 4, 2024 API access, rather than the October announcement.
- Grok 4.20 and 4.3 have only verified announcement/by-date anchors in the available official record and are contextual. GPT-6 Astra's initial selected-organization release is also contextual under the agreed public-access rule.
- The Huawei Pangu 3.0, 5.0, and 5.5 suite announcements are context milestones. Public openPangu/Pangu language weights and MaaS releases enter cadence normally.
- MiniMax abab5/5.5 and Huawei Pangu Ultra are omitted because an exact first public-access day could not be verified. Some earlier announcements establish only month-level dates. Omission means no precise cadence claim is made for those events.
- Fixed-version DeepSeek releases are included only where official notes describe a substantial capability or architecture upgrade; patch-only refreshes are excluded. A stable release is counted again only when official notes establish a material model change.
- NVIDIA's main Nemotron lineage includes both NVIDIA-pretrained models and the publicly released Llama-derived Nemotron line. These origins are disclosed in event descriptions. Secondary specialist, tooling, and minor derivative updates are excluded.
- Same-day variants are bundled into one event per vendor. Its family tags retain membership in every launched family. Family filters select whole launch events, so a bundled Opus/Sonnet event belongs to both family histories but is counted once in the vendor-wide history.
- Broad and narrower family tags can coexist (for example Claude and Claude Sonnet, or Gemini and Gemini Pro). They are overlapping subsets, never summed as independent release counts.

## Reading the cadence chart

Bars show days between consecutive **included, dated launch events**, in release-date order with equal categorical spacing. Context events are excluded, so these gaps describe the curated verified history rather than every possible launch. The first dated event has no preceding interval. Median uses completed gaps only; ongoing wait is measured to the fixed snapshot date, not the computer clock.

The timeline's date zoom affects the main plot only. Vendor cadence always uses the complete snapshot, optionally narrowed by model family. A context event can be selected for its sources without creating a cadence bar. Every count is an event count, not a count of individual model variants.

## Updating

Follow `AGENTS.md` for the complete agent-assisted refresh workflow. This file records source indexes, date precedents, corrections, and review coverage; the app does not fetch release updates itself.

## Qwen scope addition — reviewed September 30, 2026

Qwen is added at the user's request, immediately above MiniMax. The existing September 29 snapshot is retained; this is a scoped vendor addition, not an all-provider refresh. Previously supported vendor events and their coverage remain unchanged.

Reviewed Qwen's [research index](https://qwen.ai/research), paginated historical [blog](https://qwenlm.github.io/blog/), individual launch posts, and official [Qwen](https://github.com/QwenLM/Qwen), [Qwen3](https://github.com/QwenLM/Qwen3), [Qwen-VL](https://github.com/QwenLM/Qwen-VL), [Qwen3-VL](https://github.com/QwenLM/Qwen3-VL), [Qwen2.5-Omni](https://github.com/QwenLM/Qwen2.5-Omni), [Qwen3-Omni](https://github.com/QwenLM/Qwen3-Omni), and [Qwen3.8](https://github.com/QwenLM/Qwen3.8) repositories through September 29, 2026. The research index required browser rendering and loading older entries. All 42 added events embed their official evidence; same-day variants are bundled. Qwen, QwQ, QVQ, Max, VL, and Omni tags are overlapping family subsets.

Date decisions and limitations:

- Original Qwen weights use August 3, 2023; Qwen-VL uses August 22, as explicitly recorded in their official repository logs. September 25 bundles the materially retrained language weights and updated visual chat. November 30 bundles the 72B and 1.8B releases.
- Earlier launch posts explicitly document public access for Qwen1.5 (February 4, 2024), Qwen2-VL (August 29, 2024), Qwen2.5-VL (January 26, 2025), Qwen2.5-VL-32B (March 24, 2025), Qwen3.5 (February 15, 2026), Qwen3.6-35B-A3B (April 15, 2026), and Qwen3.6-27B (April 21, 2026). Some repository logs label these one or two days later; the explicitly documented earlier public launch is used. The updated Qwen1.5 post mentions later sizes; those are not backdated into its initial launch.
- Qwen2.5-Omni uses March 26, 2025 from its repository's public weights/chat log, rather than the March 27 blog date. Qwen3 uses April 29, consistent with its current official blog and repository. QVQ uses December 25, 2024 from the official visual-model release log.
- Qwen2.5-Max and Qwen3-Max-Thinking use dated public launch posts, not the dates embedded in API version names. QwQ-Max-Preview counts through public Qwen Chat. Qwen3.6-Max-Preview counts through public Studio chat; its API was described as coming soon.
- Qwen3-Max-Instruct uses September 24, 2025; an earlier public preview access date remains unverified. The post establishes material improvements in the official release, but no separate preview event is inferred. Qwen-Max-0428 and Qwen-VL-Plus are omitted pending verification of their first access dates.
- The current Qwen3.8-Max post and research index show August 3, 2026 and document public API access. An older indexed homepage shows August 2; this discrepancy remains unresolved, so the current dated launch is used. Its unchanged weights release on August 12 does not add a second cadence event. The distinct Qwen3.8-27B release uses its dated August 14 repository log.
- Qwen3.8-Flash-Next (August 26) qualifies as a public general multimodal architecture preview. Qwen3.8-Omni-Flash (September 18) qualifies through public API access and general language/multimodal agent capabilities.
- Excluded specialist Coder, Math, image, audio-only/ASR/TTS, embedding/reranking, LiveTranslate, robotics/driving/world-simulation models, benchmarks, and tooling. Routine additional sizes, quantizations, deployment updates, and unchanged re-releases are not treated as separate material launches. Initial suite launches retain their bundled variants.
- Qwen3.5-Max-Preview's March 18 Arena announcement is omitted because first direct, named public access is not established by the reviewed announcement. Qwen3-Omni-Flash's December 2025 update is omitted pending verification of first access and the material-model-change boundary; no access date is inferred from its version suffix. These gaps mean the Qwen lane is a curated history, not an exhaustive launch log.

The new hover comparison uses UTC calendar weeks starting Monday and calendar months; it compares all plotted markers, including qualified context markers, without changing cadence counts.