What Changed with Qwen3? Part 2: A History and Popularity Check for 2025 (Qwen3, Coder, Next)
After Qwen2.5 in Part 1, Qwen changed a great deal in 2025. Qwen3, released in spring, puts a “thinking mode" that reasons before answering and a “non-thinking mode" that answers right away into a single model. Qwen3-8B, a name that comes up often in local AI, has still been downloaded about 130 million times in total on Hugging Face. What changed with Qwen3 in 2025?
This article is Part 2 of a three-part series looking back at Qwen. Part 2 covers Qwen3 in 2025 and its surroundings (QwQ, Coder, Next), in the order of history, strengths, popularity, and buzz. Part 1 is here:
Why Did Qwen Become a Staple of Local AI? Part 1: From the First Qwen to Qwen2.5
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: Qwen in 2025
- 3. What was new in Qwen3?
- 4. Was it popular? The numbers
- 5. Did it get talked about? Events and reaction
- 6. Where this meets my own measurements
- 7. Discussion: where does Qwen3 stand?
- 8. What I could not find out
- 9. Summary: what Part 2 showed, and what comes next
- 10. References
Method: what I looked into and where the sources come from
I looked into three things: what Qwen released in 2025, what was new in Qwen3, and how much it was used and talked about. I confirmed release dates and specifications on Qwen’s official blog and on Hugging Face model cards. The numbers come from the public pages of Hugging Face and Ollama, and my gauge of buzz comes from Hacker News (an international tech news site).
My own measurements can be used for only part of this. For the small Qwen3 models (4B and under) and for Qwen3 speeds measured on an RTX 3090, I quote my measured values. The rest is a summary of public information. The overall lineage is summarized in this article:
The Family Tree of Open-Source LLMs: 8 Major Families
Timeline: Qwen in 2025
The number of parameters expresses a model’s size, and B means a billion. MoE (Mixture of Experts) is a design that uses only some of its experts, written in the form “total / active".
| Period | Model | Size [B] | Notes |
|---|---|---|---|
| 2025-03-05 | QwQ-32B | 32 (dense) | A model specialized for reasoning. Apache-2.0 |
| 2025-04-29 | Qwen3 | Dense: 0.6 / 1.7 / 4 / 8 / 14 / 32, MoE: 30 (3 active) / 235 (22 active) | Switching between thinking and non-thinking modes. Supports 119 languages and dialects. Trained on about 36 trillion tokens. All Apache-2.0 |
| 2025-07-21 to 08-06 | The 2507 line of Qwen3 | 30 (3 active), 235 (22 active), 4 | The thinking model and the answer-right-away model were split into separate versions (per GitHub’s update history; 08-08 was an update for one-million-token support). Context is 256K, extendable to one million tokens (model card) |
| 2025-07-22 | Qwen3-Coder | 480 (35 active); 30 (3 active) on 07-31 | For code generation and agent-style work |
| 2025-09-11 | Qwen3-Next | 80 (3 active) | An ultra-sparse MoE with a new attention mechanism. This mechanism is said to have been carried over into Qwen3.5 and later |
| 2025-09-23 to 10-21 | Qwen3-VL | 235 (22 active), 30 (3 active), 8, 4, 2, 32 | Versions that also read images. Released from the largest down (GitHub’s update history) |
What was new in Qwen3?
- Thinking and non-thinking modes can be switched within one model. The idea is to make the model think on hard problems and answer right away on easy questions, and the official side explains that you can adjust a “thinking budget".
- It released a small MoE. Qwen3-30B-A3B has 30B in total and 3B active. The official side claims it beats QwQ-32B, which has about 10 times as many active parameters (the official side’s own claim).
- Japanese became one of the officially supported languages. In the official list of 119 languages and dialects, Japanese is listed under “Other".
- Context length differs by size. In the table on the official blog, 0.6B, 1.7B, and 4B are 32K, and 8B and above are 128K. On the model cards, the standard is 32K, and 128K assumes using an extension for long contexts (YaRN).
The official side also writes that the small 4B model matches the performance of Qwen2.5-72B-Instruct. That is the official side’s claim, and I have not verified it on this blog. My own measured results for Qwen3-4B appear in a later section.
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] |
|---|---|---|---|---|
| QwQ-32B | 2025-03 | 38,999 | 3,588,693 | 2,966 |
| Qwen3-8B | 2025-04 | 9,650,573 | 130,841,100 | 2,089 |
| Qwen3-32B | 2025-04 | 3,357,796 | 59,771,021 | 755 |
| Qwen3-30B-A3B | 2025-04 | 1,320,513 | 22,039,847 | 957 |
| Qwen3-235B-A22B | 2025-04 | 485,875 | 7,369,833 | 1,111 |
| Qwen3-Coder-480B-A35B-Instruct | 2025-07 | 37,920 | 1,155,582 | 1,376 |
| Qwen3-Coder-30B-A3B-Instruct | 2025-07 | 383,753 | 15,605,514 | 1,274 |
| Qwen3-Next-80B-A3B-Instruct | 2025-09 | 249,798 | 16,605,620 | 1,059 |
| Qwen3-VL-8B-Instruct | 2025-10 | 8,752,198 | 72,669,879 | 1,177 |
On Ollama, qwen3 has been pulled 39.7M times in total, roughly the same level as qwen2.5 from Part 1 (42.2M). Qwen3-8B’s all-time total is about 130 million, lower than Part 1’s Qwen2.5-7B-Instruct (about 188 million). However, the time since release differs, so they cannot be compared simply.
Within the same size range, the numbers are skewed. Small models (8B) are pulled a lot, and huge models (235B and 480B) are pulled less. A possible background is that the larger the model, the harder it is to run on an ordinary PC. I did not confirm the reason in this investigation.
Did it get talked about? Events and reaction
| Date | Event | Hacker News points |
|---|---|---|
| 2025-03-05 | QwQ-32B (a model whose reasoning was trained with reinforcement learning) (posting date) | 480 |
| 2025-04-28 | Qwen3 (posting date) | 869 |
| 2025-07-22 | Qwen3-Coder (posting date) | 765 |
| 2025-09-12 | Qwen3-Next (posting date) | 569 |
Hacker News points are a rough gauge of how much attention overseas developers paid. Qwen3 got a bigger reaction than Part 1’s Qwen2.5-1M (307 points).
Where this meets my own measurements
I have measured the speed of small Qwen3 models and of models that run on an RTX 3090. The machines and measurement dates differ, so I show them in two tables.
| Model (Q4_K_M) | Size [GiB] | Coding tasks [out of 7] | Writing speed [tok/s] (EVO-X2 integrated GPU) |
|---|---|---|---|
| Qwen3-4B | 2.33 | 6 | 77.5 |
| Qwen3-1.7B | 1.03 | 2 | 167.8 |
| Qwen3-0.6B | 0.37 | 1 | 356.2 |
The conditions are as follows. Speed: EVO-X2, llama.cpp b11192, the median of 5 runs (September 29, 2026). Coding tasks: RTX 3090, Ollama, temperature 0, one run each. Thinking mode was left at its default (on). The smaller the model, the faster it is, and the lower its coding score.
| Model (Ollama) | Size [GB] | Writing speed [tok/s] (RTX 3090) |
|---|---|---|
| qwen3:8b | 5.2 | 129.2 |
| qwen3:14b | 9.3 | 77.1 |
| qwen3-coder:30b | 19 | 123.7 |
The RTX 3090 values come from the record of my measurements of coding-oriented models (June to July 2026, Ollama).
Discussion: where does Qwen3 stand?
Three things can be said from the published facts.
- It added a small MoE (30B-A3B) to the options that run on a home PC.
- It made thinking switchable within the model.
- Within 2025, derivatives spread (Coder, Next, VL) and led into later generations.
These are a series of facts. To claim they are the reason for its popularity, I would need to investigate the situations in which it was used.
What I could not find out
- I could not obtain third-party evaluations in Japanese.
- For the 2507 line, I have not checked the corresponding posts on the official blog. The specifications come from GitHub and the model cards.
- My own measurements differ in machine and measurement date from model to model. They are not a comparison under the same conditions.
- The download counts cannot distinguish where the downloads came from (individuals, companies, or automatic pulls from other software).
Summary: what Part 2 showed, and what comes next
In April 2025, Qwen3 released everything from small models up to 235B, together with a thinking-mode switch, all at once. Qwen3-8B has been downloaded about 130 million times in total, and Qwen3-VL-8B about 73 million.
As a next step, if you want to try it on your own PC, install `qwen3:8b` (about 5.2GB) or `qwen3:14b` (about 9.3GB) with Ollama. Part 3 covers Qwen3.5 to 3.8 in 2026: the flagship going closed, and the release of a 27B.
References
- Qwen3: Think Deeper, Act Faster (Qwen official blog)
- QwenLM/Qwen3 (GitHub, update history)
- Qwen (Hugging Face)
- qwen3 (Ollama library)
- QwQ-32B (Hacker News)
- Qwen3-Coder (Hacker News)
- Qwen (Wikipedia, secondary source)