The Chinese open-weights release timeline, dated

This is a list of dates, not an argument. Every row is a model whose weights were published for anyone to download, with a date I could trace back to the organization that published it, a parameter count as stated by that organization, and a link.

Most timelines of this kind are written from memory or from coverage, which is why they drift. I built this one from vendor pages, papers and Hugging Face’s API. This revision leaves the table at 67 rows: seven are new, one of them because an entry I dropped in the first pass turns out to be open-weights after all. The dropped list at the end has eleven entries, each with a reason.

How to read this timeline

Open weights only. A release is in the table only if the organization published downloadable weights. API-only and hosted-only models are out, including flagship ones: Qwen3-Max, the served ERNIE models, Kimi’s hosted variants, DeepSeek’s API-only models.

A date I could trace. The Date column has two kinds of value. A plain date is the date on the organization’s own dated page — a research blog post, a news post, a release page, an accepted paper. A date marked † is the Hugging Face repository creation timestamp in UTC, read from the Hugging Face API on 2026-10-09, used when the vendor had no dated page I could read. The two can differ by a few days, and that is usually preparation rather than a different release: DeepSeek-V3’s weights repository was created 2024-12-25 12:52 UTC and the announcement is dated 2024-12-26.

Parameter counts come from the vendor, or they are blank. Total parameters with activated parameters for mixture-of-experts models, as stated on the announcement, the model card, or the repository holding the same weights. A family that ships many sizes gets a range. Where none of the vendor’s pages for that release states a count, the cell is — rather than a number carried over from a sibling model, even when the lineage makes the number obvious.

One link per row, to the vendor or the weight repository. No aggregator links, no news coverage. That makes most rows single-source by construction: each cites the vendor’s own statement about its own release, and every † row is single-source by definition.

Scope. Language models and the flagship multimodal lines that carry the same brand and weights: Qwen3-VL, Qwen3-Omni, MiniMax-M3. Pure image, video and speech models are out of scope.

When the vendor’s page is client-rendered

Two of the labs whose releases I date most often publish announcements as a page that ships no text to a plain HTTP request. z.ai’s blog returns a 598-byte HTML shell that loads a JavaScript bundle; qwen.ai’s blog returns a shell that fetches its content after load. Both render fine in a browser, which is what a reader clicking the link sees.

For z.ai I did read something: the date string sits inside the page’s own bundle, so the bundle is where I checked GLM-4.5 (2025-07-28), GLM-4.6 (2025-09-30), GLM-4.7 (2025-12-22) and GLM-5 (2026-02-12). That is weaker evidence than a rendered page, and it is why GLM-5.2 stays out: its bundle contains 2026-06-16 and no other date, and a date visible only inside a build artifact is not one I will print as sourced.

Qwen’s older blog carries the full text and a date; the newer qwen.ai copies do not. So the three Qwen3-* rows dated from qwen.ai in the first pass now cite the weight repository and carry a †.

2023

Q1–Q2

Q3–Q4

Baichuan 2 is the only 2023 row dated by a paper: arXiv:2309.10305 is dated 2023-09-19, while its weight repositories appear earlier, on 2023-08-30 †, so the paper is the single source for that date. The two ChatGLM cells are blank because their cards do not state a count: 6.2 billion is on the ChatGLM-6B card and not on the cards of its two successors, and copying it across is the habit that makes timelines of this kind drift.

2024

Q1

Qwen1.5 was in the dropped list in the first pass, because I looked for it on the new blog and found nothing readable. The old blog still carries the post, dated 2024-02-04, listing six sizes plus 110B.

Q2–Q3

Q4

The DeepSeek point releases inside a model line are dated on DeepSeek’s own changelog, one page per release. Hunyuan-Large’s 389B/52B comes from arXiv:2411.02265, submitted 2024-11-04, three weeks after the weights appeared.

2025

Q1

The two DeepSeek updates here do not restate a parameter count on their own changelog pages; the cells carry the count from the repositories that hold those weights, which are the V3 and R1 repositories listed above.

Q2

Q3

Q4

2026

Q1

Q2

The DeepSeek-V4 row now carries both counts; the Flash model’s 284B total is stated on the same announcement page as the Pro figures. Step 3.7 Flash moved to a † date: its GitHub repository has no releases, and the page I first cited carries a creation timestamp rather than a release date.

Q3

Hy3 is the entry whose status changed in this revision. In the first pass I dropped it because the page I found did not say whether weights were published. It does: Hy3 is open-sourced under Apache 2.0 with weights on Hugging Face, and Tencent’s newsroom names the parameter counts. The V4-Pro GA row repeats the count from the April preview announcement, because the GA page does not restate it. The Qwen3.8 row is dated by the repository of the 27B model, the first of the line to appear; the flagship 2.4T-A95B repository was created three days later, and Qwen’s own announcement page is one of the client-rendered ones described above.

Two rows elsewhere disagree with third-party timelines for the same reason: LongCat-2.0’s repository was created 2026-07-05 while secondary coverage puts its launch at the end of June, and DeepSeek-V4-Flash’s repository was created 2026-04-22, two days before the announcement page.

What changed in this revision

Seven rows were added and three cells changed, listed because the edit history is the point of a table like this:

1. Qwen1.5 promoted into 2024 Q1 (2024-02-04), after finding the dated post on the old Qwen blog.
2. Hunyuan Hy3 promoted into 2026 Q3 (2026-07-06). It was dropped in error.
3. Four DeepSeek point releases added: V2.5-1210 (2024-12-10), V3-0324 (2025-03-25), R1-0528 (2025-05-28) and V4-Pro GA (2026-08-13). All four are dated on DeepSeek’s own changelog, the same source as V3.2 and V3.2-Exp.
4. Qwen3.8 added (2026-08-05 †), which the first pass omitted entirely.
5. Two ChatGLM parameter cells blanked. The ChatGLM2-6B and ChatGLM3-6B cards do not state a count, so the cells are now —.
6. DeepSeek-V4-Flash gained its total. The row said “Flash 13B active” and dropped the 284B total, which the same page states.

Three rows changed source or date type for the same structural reason: the Qwen3-Next, Qwen3-Omni and Step 3.7 Flash rows now cite a weight repository with a †, because the pages cited in the first pass were either unreadable to a plain request or carried no release date.

How this table gets updated

The table is edited in place. Every quarter I re-read the vendor pages for the labs already listed, then check Hugging Face for new repositories under those organizations. A release enters the table when it has a date and a link; until then it sits in the dropped list, with the reason.

Two rules keep it stable: I do not replace a † date with a later one unless the vendor publishes its own dated page, and I do not remove a superseded row.

The current table has 67 entries. Twenty-three of them fall in the twelve months from October 2025 to October 2026. That is a count of this table and nothing more; it is not a claim about the field, because the table only holds releases I could date.

What I could not check

Eleven candidates were dropped, for these reasons:

1. Yi-1.5 (May 2024) — the weight repository (01-ai/Yi-1.5-34B) was created 2024-05-11 †, and I could not date the announcement itself.
2. InternLM2 and InternLM2.5 — the weight repositories (internlm/internlm2-7b, internlm/internlm2_5-7b-chat) look like renamed predecessors, so the creation timestamp does not mark the release.
3. MiniCPM 1 and 2 — same problem; the repository (openbmb/MiniCPM-2B-sft-bf16) is now a monorepo for a whole series.
4. Baichuan 3 and 4 — I could not find weight artifacts at all.
5. ERNIE 5.0 and ERNIE 5.1 — Baidu’s blog dates them 2026-02-06 and 2026-05-09 (ERNIE 5.0, ERNIE 5.1) and describes ERNIE 5.0 as a 2.4-trillion-parameter model, but I found no open-weight repository for either.
6. Kimi K3 — the model card (moonshotai/Kimi-K3) carries the full architecture table (2.8T total, 104B activated) and a license, but no release date. The repository timestamp is 2026-06-13 and secondary coverage says mid-July. I could not reconcile that.
7. Kimi K2.5 — same shape of problem: the weight repository timestamp (moonshotai/Kimi-K2.5) is 2026-01-01 while the launch is reported at the end of January.
8. GLM-5.2 — the post is client-rendered and the only date string in its bundle is 2026-06-16. I am not willing to print that as the release date.
9. DeepSeek-V3.1-Terminus — a mid-cycle rename with a 2025-09-22 † repository (deepseek-ai/DeepSeek-V3.1-Terminus); including it would have double-counted the V3.1 line.
10. Qwen-Image-2.1 — a dated weight repository (Qwen/Qwen-Image-2.1, 2026-09-14 †) but an image model, outside this table’s scope by the rule stated at the top.
11. The Step-Audio line — the same, for speech. The repositories exist (stepfun-ai/Step-Audio-R1.1) and I did not date them, because they would need their own table with its own scope.

Three other limits on this table:

• The † timestamps are preparation times, not launch times. A repository can be created days before the announcement, and from outside I cannot see whether weights were uploaded at once or in pieces. Where the gap is large I noted it above the table.
• The dates are the vendor’s own. No release date here has been confirmed by a second source, and one link per release is deliberate: a vendor correcting itself is the signal worth keeping, and coverage repeating the vendor is not a second source.
• License differences are not reflected. Apache-2.0, MIT and the vendor-specific licenses Kimi, MiniMax and LongCat ship under are not the same permission. Each row links to the page that states its license, and for Hy3 that page states Apache 2.0.