Why Did Qwen Become a Staple of Local AI? Part 1: From the First Qwen to Qwen2.5, a History and Popularity Check (2023-2024)
If you try to run AI locally, the name “Qwen" always shows up in the list of models. It is a model series made by Alibaba, and on Ollama, Qwen2.5, released in 2024, has been pulled 42.2M times in total. Yet the very first model was a small 7B one. How did Qwen get to the point of being called a staple of local AI?
This article is Part 1 of a three-part series looking back at Qwen. Part 1 covers the period from the first release to Qwen2.5 (2023 to 2024). I looked into Qwen’s history, what it was good at, and whether it was popular, using publicly available information.
This reflects the state of things as of October 12, 2026.
- 1. Method: what I looked into and where the sources come from
- 2. Timeline: from the first Qwen to Qwen2.5
- 3. What was Qwen good at?
- 4. Was it popular? The numbers
- 5. Did it get talked about? Events and reaction
- 6. Discussion: why did Qwen2.5 become a staple?
- 7. What I could not find out
- 8. Summary: what Part 1 showed, and what comes next
- 9. References
Method: what I looked into and where the sources come from
I looked into three things: in what order Qwen released which models, how the license (the conditions for using a model) changed, and how much the models have been used. I checked the numbers in five places: Qwen’s official blog, Hugging Face (a site that distributes models), Ollama (a tool for running AI locally), GitHub, and Hacker News (an international tech news site).
I could not find any record of this blog measuring models older than Qwen2.5, so this installment is an investigation with no measurements of my own. In Part 3, I will use numbers I measured for the newer generations. The overall lineage is summarized in this article:
The Family Tree of Open-Source LLMs: 8 Major Families
I confirmed release dates on Qwen’s official blog (dates in China time). The dates on the official blog can differ by a day from GitHub’s update history or from the creation date of a Hugging Face repository.
Timeline: from the first Qwen to Qwen2.5
The number of parameters expresses a model’s size, and B means a billion. “Dense" means a design that uses all of its parts every time, and “MoE" (Mixture of Experts) means a design that uses only some of its parts, called “experts".
| Generation | Release | Size [B] | License | Notes |
|---|---|---|---|---|
| Qwen (first generation) | 2023-08 (7B), 2023-09 (14B), 2023-11-30 (72B and 1.8B) | 1.8 / 7 / 14 / 72 (dense) | Tongyi Qianwen License (a research-only version for 1.8B) | The 7B was trained on over 2.4 trillion tokens (per the model card) |
| Qwen1.5 | 2024-02-04 | 0.5 to 72 (dense). 32 and 110 added in 2024-04 | Tongyi Qianwen License (a research-only version for some small models) | The first generation to release everything from 0.5B to 72B at once. First MoE (Qwen1.5-MoE-A2.7B, 2024-03-28) |
| Qwen2 | 2024-06-07 | 0.5 / 1.5 / 7 / 72 (dense), 57 (MoE, 14 active) | Apache-2.0 except 72B; 72B uses the Tongyi Qianwen License | Everything except the 72B moved toward open terms |
| Qwen2.5 | 2024-09-19 | 0.5 to 72 (dense) | Mostly Apache-2.0 (3B uses the research-only Qwen Research License, 72B uses the Qwen License) | Trained on 18 trillion tokens. Context is 128K for 7B and up, and 32K for 0.5 to 3B. Followed by Coder (1.5B and 7B in 2024-09, six sizes from 0.5B to 32B on 2024-11-12), VL (2025-01-26), and a 1M-context version (2025-01-27) |
“Apache-2.0" is a license that gives a lot of freedom, including commercial use. The fact that Qwen2 (except the 72B) and most of Qwen2.5 came under these terms probably contributed to how easy they were to use. However, licenses differ by model, and some are limited to research, so you need to check each model’s card before using one.
What was Qwen good at?
- From Qwen1.5 on, it released sizes from 0.5B to 72B all at once. Being able to choose a size that fits your own PC is a big advantage in local AI.
- For code, Qwen2.5-Coder came out in 1.5B and 7B in September 2024, and in six sizes including 32B in November. The 32B version is still downloaded about 800,000 times in the last 30 days on Hugging Face.
- Qwen2.5-VL, a version that can also read images and came out in January 2025, was downloaded about 5.28 million times in the last 30 days in its 7B version.
- For long documents, it released Qwen2.5-1M with a one-million-token context in January 2025, which drew 307 points on Hacker News (posted 2025-01-26).
Assessments of strengths are a place where official claims and third-party measurements easily get mixed. What I list here are facts you can confirm on public pages: sizes, how the models are offered, and download counts. I could not obtain a third-party evaluation of its Japanese performance. The Ollama description of Qwen2.5 lists Japanese among its 29 supported languages.
Was it popular? The numbers
The Hugging Face download counts are the last-30-days value shown as “Downloads last month" on a model page, and the all-time total available through the API. I retrieved them on October 12, 2026.
| Model | Released | Downloads [times, last 30 days] | Downloads [times, all time] | Likes [count] |
|---|---|---|---|---|
| Qwen-7B (first generation) | 2023-08 | 68,948 | 2,225,816 | 399 |
| Qwen1.5-7B | 2024-02 | 73,390 | 3,619,238 | 56 |
| Qwen2-7B-Instruct | 2024-06 | 287,651 | 26,332,990 | 688 |
| Qwen2.5-7B-Instruct | 2024-09 | 7,753,092 | 187,956,173 | 2,418 |
| Qwen2.5-Coder-32B-Instruct | 2024-11 | 800,328 | 13,545,741 | 2,167 |
| Qwen2.5-VL-7B-Instruct | 2025-01 | 5,277,706 | 104,161,681 | 1,735 |
Comparing the 7B line, download counts grow with each generation, and Qwen2.5-7B-Instruct stands out. Its total is about 188 million. I did not check in this investigation why it is still pulled about 7.75 million times a month, two years after release.
On Ollama, qwen2.5 has been pulled 42.2M times in total, with 133 tags (differences in size and quantization) published. Besides that, qwen2.5-coder has 22.3M, qwen2 has 6.3M, and the original qwen has 8M. Qwen3 has 39.7M, but it came out about 7 months after Qwen2.5, so the difference in time on the market needs to be taken into account when comparing.
Did it get talked about? Events and reaction
| Date | Event | Hacker News points |
|---|---|---|
| 2024-11-27 | QwQ-32B-Preview (an o1-style reasoning model) appears (posting date) | 438 |
| 2025-01-26 | Qwen2.5-1M (one-million-token context) (posting date) | 307 |
Hacker News points are a rough gauge of how much attention overseas developers paid. QwQ got 438 points and Qwen2.5-1M got 307. Points alone do not tell you the volume of press coverage or general public interest.
Discussion: why did Qwen2.5 become a staple?
Three things can be said from the published facts.
- The range of sizes was wide, and from Qwen2 on, most sizes came under open licenses.
- A code version and an image version were released as models of the same generation.
- They became usable in tools such as Ollama, and they are still being downloaded today.
These are a series of facts that point to reasons for the popularity. I did not confirm in this investigation a causal claim that “this is exactly why it became popular".
What I could not find out
- I did not look into third-party evaluations in Japanese, or comparisons with other companies’ models at the time.
- The popularity numbers cannot distinguish where the downloads came from (individuals, companies, or automatic pulls from other software).
- The license column summarizes each model’s conditions. Before using a model, you need to check the full text on its model card.
- The dates on Qwen’s official blog are in China time, so they can differ from Japan time by a day.
Summary: what Part 1 showed, and what comes next
Qwen started with a 7B model in August 2023. With Qwen2.5 it assembled sizes from small up to 72B, plus code and image versions, and became a staple of local AI. Qwen2.5-7B-Instruct is still downloaded about 7.75 million times in the last 30 days or so.
If you want to try it on your own PC now, installing `qwen2.5:7b` (about 4.7GB) with Ollama first will give you a feel for the staple. Part 2 covers Qwen3 in 2025: the switch that turns thinking mode on and off, and the arrival of an MoE (30B-A3B) that runs even on a home PC.
References
- Qwen official blog
- QwenLM/Qwen3 (GitHub, update history)
- Qwen (Hugging Face)
- qwen2.5 (Ollama library)
- Qwen2.5-1M (Hacker News)
- QwQ (Hacker News)
- Qwen (Wikipedia, secondary source)