Skip to content
FrontierDot

Qwen model timeline: Qwen-7B to Qwen3-Coder, licenses (2023–2025)

Live timeline · last updated · 1 entry in the update log

Qwen’s language-model line is long, and the license is not one grant. A 72B model can sit under a different text from the 7B model released beside it. This page uses the Qwen Team blog for dates and says when a table date and a “we now release” sentence are not the same day.

As of October 21, 2026 the latest language-model post we could date on the Qwen Team blog is Qwen3-Coder, on July 22, 2025. Names such as Qwen3.5 and later were not dated posts on that blog. Image models are covered in separate Front posts.

Read on 2026-10-21: the Qwen Team posts Introducing Qwen, Qwen1.5, Qwen2, Qwen2.5, Qwen2.5-Coder, Qwen2.5-1M, Qwen2.5-Max, Qwen3 and Qwen3-Coder, plus the Hugging Face cards for Qwen3-235B-A22B and Qwen3-Coder-480B-A35B-Instruct. “First” and “state of the art” sentences are Qwen’s. We did not run the models. qwen.ai/blog is a client-side shell, so it is not a source.

Timeline

  1. The Qwen Team post published January 23, 2024 includes a table that dates Qwen-7B to 23.08.03. The row gives a 32K context, system prompt not supported, and 2.4 trillion pretrained tokens. The post calls Qwen-7B the starting point of the open-source line. We did not find a separate August 3, 2023 announcement page in this pass, so the node date is the table date and the source is the January post that prints the table. The same post’s phrase “we now release” is about the January write-up, not a second claim that 7B shipped in January.

    Qwen Team: Introducing Qwen

  2. The same January 23, 2024 table dates Qwen-14B to 23.09.25, with an 8K context, system prompt not supported, and 3.0 trillion pretrained tokens. As with Qwen-7B, the source page is later than the table date. We are not promoting the blog’s publication day into a release day for this row. Eight thousand tokens of context is the figure in the table. Later Qwen posts extend context. This row does not.

    Qwen Team: Introducing Qwen

  3. The same table dates both Qwen-1.8B and Qwen-72B to 23.11.30. Both rows show a 32K context and system prompt supported. Pretrained tokens are 2.2 trillion for the 1.8B model and 3.0 trillion for the 72B model. The January 23, 2024 prose says the team is now releasing Qwen-1.8B, Qwen-7B, Qwen-14B and Qwen-72B. That sentence and the November 30 table cells are both on the page. The node uses the table date. It does not pretend the prose and the table name one day.

    Qwen Team: Introducing Qwen

  4. Qwen Team open-sources Qwen1.5 base and chat models. The page lists 0.5B, 1.8B, 4B, 7B, 14B, 32B, 72B and 110B, plus a mixture-of-experts model. We do not have a separate announcement body for the 110B checkpoint, so this node does not claim that every size on that list shipped on February 4. The fact we will stand on is the list itself: base and chat, at the sizes the page names, on this date.

    Qwen Team: Qwen1.5

  5. Qwen Team posts Qwen2. The visible date on the page is June 7, 2024. An HTML dateModified of July 16, 2024 is not the day we use. Sizes are Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, Qwen2-57B-A14B and Qwen2-72B. The post says the models were trained on data in 27 additional languages. Context in the size table is 32K, 32K, 128K, 64K and 128K for those five sizes in that order. A separate sentence says context was extended to 128K for Qwen2-7B-Instruct and Qwen2-72B-Instruct. The table and that sentence are not the same claim: one is a five-row table, the other names two instruct models. The license sentence is quoted in the license section. Qwen2-72B stays on the original Qianwen License. The smaller dense models and the 57B-A14B model move to Apache 2.0.

    Qwen Team: Qwen2

  6. Qwen Team open-sources Qwen2.5. The post says most models support a context of 128K tokens, which it also writes as 131,072, and can generate up to 8K tokens. The majority are Apache 2.0. Qwen2.5-3B is under the Qwen Research License. Qwen2.5-72B is under the Qwen License. The table snippet we used gives Qwen2.5-0.5B as 0.49 billion parameters, 32K context, 8K generation, Apache 2.0; Qwen2.5-1.5B as 1.54 billion with the same context, generation and license; and Qwen2.5-3B as 3.09 billion under the Qwen Research License. The post also says the team is open-sourcing Qwen2.5-3B, 14B and 32B, and offering Qwen-Plus and Qwen-Turbo through Alibaba Cloud Model Studio. Sibling posts that day for Coder and Math are titles on the blog index. Their bodies are not in this node.

    Qwen Team: Qwen2.5

  7. Qwen Team open-sources the Qwen2.5-Coder series. The post says Qwen2.5-Coder-32B-Instruct is the current state-of-the-art open-source code model. That ranking is Qwen’s claim. We did not rerun the code benchmarks, and we do not treat the sentence as a measured result from this site. The dated fact is the open-source release of the coder series on this day, with 32B-Instruct as the checkpoint Qwen points at.

    Qwen Team: Qwen2.5-Coder

  8. Qwen Team open-sources Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, with context up to 1 million tokens. The post calls this the first time such a context has been opened this way. “First” is Qwen’s wording. The checkpoints named in the post are the two instruct models above, not the entire Qwen2.5 size list from September.

    Qwen Team: Qwen2.5-1M

  9. Qwen Team posts Qwen2.5-Max. The page’s published time is January 28, 2025, 23:00 in China, which is still January 28 in China and 15:00 UTC. We use January 28. The model is offered through Alibaba Cloud and Qwen Chat. The post we read does not say the weights are open. This node is an API release, not an open-weight release, and it is a different artifact from the 1 million token instruct models posted the day before.

    Qwen Team: Qwen2.5-Max

  10. Qwen Team posts Qwen3. The visible date is April 29, 2025. The published time on the page is April 29, 2025, 04:00 in China. We do not use an older snippet that said April 28. The flagships are Qwen3-235B-A22B, 235 billion total and 22 billion activated, and Qwen3-30B-A3B, 30 billion total and 3 billion activated. Dense open-weight models under Apache 2.0, in the blog’s sentence, are Qwen3-32B, 14B, 8B, 4B, 1.7B and 0.6B. That Apache sentence is attached to the dense list. It does not, by itself, license the mixture-of-experts models. The Hugging Face card for Qwen3-235B-A22B lists apache-2.0, which is a second source, stated in the license section. Dense context on the blog is 32K for 0.6B, 1.7B and 4B, and 128K for 8B, 14B and 32B.

    Qwen Team: Qwen3

  11. Qwen Team posts Qwen3-Coder. The published time is July 22, 2025, 21:00 in China. Qwen3-Coder-480B-A35B-Instruct is a 480 billion parameter mixture-of-experts model with 35 billion active parameters. The post says it supports 256K tokens natively and 1 million tokens with extrapolation methods. The Hugging Face card Qwen/Qwen3-Coder-480B-A35B-Instruct lists the license apache-2.0. Native context and extrapolated context are different numbers. Quoting only 1 million leaves out the limit the post calls native.

    Qwen Team: Qwen3-Coder

How we date a Qwen release

The source we can actually read is the Qwen Team blog on qwenlm.github.io, including its RSS feed for titles and timestamps. qwen.ai/blog did not return article text. A node date is either the date printed on the post or, for the original Qwen family, the date in the table of the January 23, 2024 post. When those differ, the node says so.

An HTML dateModified is not a release. Qwen2 shows June 7, 2024 on the page and a later dateModified. We use June 7. Qwen2.5-Max is dated from its published time, January 28, 2025, 23:00 in China, which remains January 28 in UTC. Qwen3’s published time is April 29, 2025, 04:00 in China. We do not move it to April 28.

The last check is October 21, 2026. The blog index’s latest language-model posts we are willing to name stop at Qwen3-Coder on July 22, 2025. The index also lists Qwen3Guard on September 23, 2025. We did not read that article, so it is not a node. Later names that show up on third-party trackers, including Qwen3.5, Qwen3.6, Qwen3.7 and Qwen3.8, have no dated Qwen Team post in the sources we read.

Which license applies to which size

Qwen2 is the clearest split. The post says that Qwen2-72B and its instruction-tuned models still use the original Qianwen License, and that Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B and Qwen2-57B-A14B adopt Apache 2.0. A download page that says “Qwen2” without the size is not a license.

Qwen2.5 repeats the pattern with three names. Most models are Apache 2.0. Qwen2.5-3B is under the Qwen Research License. Qwen2.5-72B is under the Qwen License. The small-end table matches that: 0.49 billion and 1.54 billion are Apache 2.0, and 3.09 billion is the research license. The 0.5B and 1.5B marketing names and the 0.49 and 1.54 table figures are both on the post. We print both rather than rounding one away.

Qwen3’s blog attaches Apache 2.0 to the dense models: 32B, 14B, 8B, 4B, 1.7B and 0.6B. The mixture-of-experts flagships are described as open weight in the post. The Apache sentence we are using does not name them. Separately, the Hugging Face card for Qwen3-235B-A22B lists apache-2.0. Cite the card for that checkpoint’s license, and cite the blog for the dense list. Qwen3-Coder-480B-A35B-Instruct also lists apache-2.0 on its card.

Context lengths, without collapsing the table

The original table is short and uneven. Qwen-7B is 32K, Qwen-14B is 8K, and the November 30 models are 32K. Qwen2’s five-row table is 32K, 32K, 128K, 64K and 128K for 0.5B, 1.5B, 7B, 57B-A14B and 72B. The instruct sentence about 128K names only Qwen2-7B-Instruct and Qwen2-72B-Instruct. The 57B-A14B row at 64K is easy to lose if you remember only the headline.

Qwen2.5 says most models support 128K, written as 131,072 tokens, and generation up to 8K, while the small rows we quoted are 32K context and 8K generation. “Most” is the post’s word. It is not “all.” Qwen2.5-1M then opens two instruct checkpoints, 7B and 14B, at up to 1 million tokens. Qwen2.5-Max, the next day, is an API model and is not that weight release.

Qwen3 dense context is 32K for 0.6B, 1.7B and 4B, and 128K for 8B, 14B and 32B. The blog’s mixture-of-experts table gives Qwen3-30B-A3B as 48 layers, 32 query heads and 4 key-value heads, 128 experts with 8 active, and 128K context. Qwen3-235B-A22B is 94 layers, 64 query heads and 4 key-value heads, 128 experts with 8 active, and 128K context. Qwen3-Coder is the outlier on length: 256K native, and 1 million with extrapolation. Those are Qwen’s figures for the architecture they published.

What the blog has not dated

  • Qwen3.5, Qwen3.6, Qwen3.7 and Qwen3.8. Not dated posts on the Qwen Team blog we checked on October 21, 2026.
  • Qwen3Guard, listed on the index for September 23, 2025. The article body was not read, so there is no node.
  • Qwen1.5-110B as a February 4, 2024 release. The February 4 page lists 110B. We do not have a separate body that dates that size.
  • Image models. Front already covers Qwen Image 2.1 Turbo and Qwen Image 2.1 transparency and edits. This page does not add an image node from a title alone.
  • Qwen-Plus and Qwen-Turbo prices. The September 19, 2024 post says they are available through Alibaba Cloud Model Studio. It does not, in the sentences we used, give a rate card.

Related timelines

The same format covers Claude, OpenAI, Gemini, Grok and Meta Muse. Labs covered in this batch are Moonshot Kimi, Mistral and DeepSeek. Image checkpoints are in Qwen Image 2.1 Turbo and Qwen Image 2.1 transparency and edits. The Front blog index lists every post.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

What is the latest Qwen language model on this timeline?

Qwen3-Coder, posted July 22, 2025. As of October 21, 2026 we did not find a dated Qwen Team post for Qwen3.5 or any later language model.

Is every Qwen model Apache 2.0?

No. Qwen2-72B stays on the Qianwen License while smaller Qwen2 models move to Apache 2.0. Qwen2.5-3B uses the Qwen Research License and Qwen2.5-72B uses the Qwen License. Qwen3’s dense list is Apache 2.0 on the blog. The 235B mixture-of-experts card on Hugging Face lists apache-2.0 separately.

When was the first Qwen model released?

The January 23, 2024 Qwen Team post’s table dates Qwen-7B to August 3, 2023. We use that table date. We did not find a separate August 2023 announcement page.

Is Qwen2.5-Max open weight?

The January 28, 2025 post describes access through Alibaba Cloud and Qwen Chat. It does not say the weights are open. The open 1 million token checkpoints are the January 27 instruct models.

Update log

  • — First published. Checked Qwen Team blog posts from the January 2024 Qwen introduction through Qwen3-Coder, as of October 21, 2026.

Sources