Why Did Qwen Reopen Its Weights After Closing the Flagship? Part 3: The History and Popularity of Qwen3.5 to 3.8 in 2026
In 2026, Qwen stopped releasing the weights of its flagship with Qwen3.7. Then in August 2026, with Qwen3.8, it released the weights of both: the 2.4T flagship under a custom license, and the 27B, which fits on a home GPU, under Apache-2.0. On Hugging Face, Qwen3.8-27B has been downloaded about 7.12 million times in the last 30 days. Why did a company that had once closed its flagship open the weights again? What has changed about Qwen over the past year?
This article is Part 3 of a three-part series looking back at Qwen. Part 3 covers Qwen3.5 to Qwen3.8 in 2026, in the order of history, strengths, popularity, and buzz. Part 2 is this article:
What Changed with Qwen3? Part 2: The History and Popularity of 2025 (Qwen3, Coder, Next)
This reflects the state of things as of October 14, 2026.
- 1. Method: scope and sources
- 2. Timeline: Qwen in 2026
- 3. What was new in Qwen3.5 to 3.8
- 4. Why did it open the weights again after closing the flagship?
- 5. Was it popular? The numbers
- 6. Was it talked about? Events and reactions
- 7. Connection to this blog’s measurements
- 8. Discussion: how has Qwen changed?
- 9. Limits: what I could not confirm
- 10. Summary: what Part 3 showed, and the end of the series
- 11. References
Method: scope and sources
I looked into three things: what Qwen released in 2026; how the flagship weights stopped being released with Qwen3.7, and how they were released again in August 2026; and how much the models were used and talked about.
- Release dates and specs: I checked the update history in GitHub READMEs, model cards on Hugging Face (a site that distributes models), and Wikipedia (a secondary source). I could not read the body of Qwen’s official blog (qwen.ai) in this environment, so for some generations I have not confirmed the official release date.
- Numbers: download counts come from public Hugging Face pages (retrieved on October 14, 2026), and the measure of buzz is the score on Hacker News (an international tech news site).
- This blog’s measurements: I draw on the speeds of Qwen3.8-27B, small Qwen3.5 models, and Qwen3.8-Flash-Next. I took no new measurements for this article.
The overall lineage is summarized in this article:
The Family Tree of Open-Source LLMs: 8 Major Families
Timeline: Qwen in 2026
The number of parameters expresses a model’s size, where B means a billion and T means a trillion. MoE (Mixture of Experts) is a design that uses only some of its parts, called “experts", and is written as “total / active". Because I could not read the official blog, the dates are given by Hacker News post date, GitHub entry, or Hugging Face creation date, as noted in each row.
| Date | Model | Size [B] | Notes |
|---|---|---|---|
| 2026-02-03 (HN) | Qwen3-Coder-Next | 80 (3 active) | For coding. Apache-2.0. Context of 262,144 tokens (per the Hugging Face card) |
| 2026-02-16 (HN) | Qwen3.5 | 397 (17 active), 122 (10 active), 35 (3 active), 27; small ones 0.8 / 2 / 4 / 9 | Apache-2.0. Every size can also read images and video. The hybrid attention design (Gated DeltaNet), introduced in Qwen3-Next (September 2025), became standard across all sizes. The Hugging Face creation date for 27B and 35B-A3B is 02-24 |
| 2026-03 | Head of Qwen steps down | − | Wikipedia records that Lin Junyang, the head of Qwen, stepped down, and that Alibaba stated it would continue releasing models |
| 2026-04-16 to 22 (HN) | Qwen3.6 | 35 (3 active), 27 | Apache-2.0. Context of 262,144 tokens (1,010,000 with extension). Features “thinking preservation" (per the Hugging Face card) |
| 2026-05 to 06 | Qwen3.7 | − | Wikipedia’s description only. The flagship is said to have been offered through an API, with the weights not released |
| 2026-08-03 (HN) | Qwen3.8-Max | about 2.4T | Generally available in the cloud (news reports) |
| 2026-08-12 (HN; Hugging Face creation date 08-08) | Qwen3.8-2.4T-A95B | 2.4T (95 active) | Weights released. A text-only model (no image input, thinking cannot be turned off). Custom license (Qwen3.8-Max License) |
| 2026-08-14 (HN; Hugging Face creation date 08-05) | Qwen3.8-27B | 27 (dense) | Apache-2.0. Context of 262,144 tokens (1 million with extension). Supports image and video input |
| 2026-08-26 (GitHub) | Qwen3.8-Flash-Next | 125 (6 active) + 51 of N-gram embeddings | The GitHub description says it is an early release of the design to be used in the next generation (Qwen4). Custom Qwen Community license |
What was new in Qwen3.5 to 3.8
- The architecture became a hybrid. The design that swaps in a different attention mechanism (a Gated DeltaNet hybrid) arrived with Qwen3-Next (September 2025) and became standard across all sizes (0.8B to 397B) in Qwen3.5 (per the Hugging Face card). Qwen3.8-27B is a dense model with 64 layers.
- Images and video can be read from the start. The Qwen3.5 card describes video input, and Qwen3.8-27B also accepts images and video. The 2.4T flagship, however, is a text-only model with no image input.
- Long context became standard. 262,144 tokens is the standard from Qwen3.5 onward (about 1 million with extension).
- You can choose how hard the model thinks. The Qwen3.8-27B card says thinking is on by default and that the effort can be set (reasoning_effort).
The Qwen3.8-27B card claims that it beats the larger Qwen3.7-Plus on coding and office-work evaluations. Numbers such as SWE-bench Pro 61.7 are listed, but these are the vendor’s own claims and depend on the evaluation conditions. I have not verified them on this blog.
Why did it open the weights again after closing the flagship?
Here are the events of 2026 in order.
- In March 2026, the head’s departure drew attention on Hacker News (14 points). Concerns were raised that the release policy might change, but Wikipedia says Alibaba stated it would continue releasing models.
- From spring to summer 2026, the Qwen3.7 flagship was offered through an API. The weights were not released (per Wikipedia; I could not confirm a primary source).
- In August 2026, the weights of two models, the 2.4T flagship (Qwen3.8-2.4T-A95B) and the 27B, were released almost at the same time. The 2.4T is under a custom license and the 27B under Apache-2.0. Flash-Next was also released on the 26th of the same month.
I checked the 2.4T license against the original LICENSE text on Hugging Face. There are two conditions. The first applies when the user (and its affiliates) runs a “Model as a Service" or “AI Work Assistant" business with total revenue above 50 million US dollars over 12 consecutive months: a separate license is needed before commercial use. Internal use that is not offered outside the company is excluded. The second applies to commercial products with more than 100 million monthly users or more than 20 million US dollars in monthly revenue: the model name must be shown. Please read the full LICENSE before using it.
I could not find an official explanation of why the weights were released again. From this sequence of facts alone, I cannot give a definite answer to the question in this article’s title. What I could confirm is only this: the Qwen3.7 flagship was not released, and with Qwen3.8 the 2.4T was released under a custom license and the 27B under Apache-2.0.
Was it popular? The numbers
The Hugging Face download counts are the last-30-days value shown as “Downloads last month" on each model page, and the cumulative value available through the API. I retrieved them on October 14, 2026.
| Model | Released | Downloads [count, last 30 days] | Downloads [count, cumulative] | Likes [count] |
|---|---|---|---|---|
| Qwen3-Coder-Next | 2026-02 | 592,172 | 7,122,948 | 1,693 |
| Qwen3.5-397B-A17B | 2026-02 | 521,542 | 5,855,490 | 1,572 |
| Qwen3.5-35B-A3B | 2026-02 | 1,474,849 | 20,382,462 | 1,532 |
| Qwen3.5-27B | 2026-02 | 1,872,253 | 20,817,120 | 1,066 |
| Qwen3.6-35B-A3B | 2026-04 | 3,649,438 | 31,200,238 | 2,980 |
| Qwen3.6-27B | 2026-04 | 2,260,408 | 28,638,698 | 2,328 |
| Qwen3.8-27B | 2026-08 | 7,119,569 | 15,084,605 | 17,637 |
| Qwen3.8-2.4T-A95B | 2026-08 | 32,860 | 93,117 | 1,309 |
| Qwen3.8-Flash-Next | 2026-08 | 2,065,917 | 2,711,741 | 6,177 |
Two points stand out. The first is Qwen3.8-27B: only about two months after release, it has about 7.12 million downloads in the last 30 days, more than the roughly 2.26 million to 3.65 million of the two Qwen3.6 models. Its 17,637 likes are also an order of magnitude more than the other models. The second is the 2.4T flagship, which has only about 33,000 downloads in the last 30 days. Its huge size, which makes it hard to run on a home PC, is a possible background, but I have not confirmed this in this investigation.
Ollama pull counts are about 3.7 million for qwen3.8 (October 14, 2026) and about 7.1 million for qwen3.6 (October 12). The retrieval dates differ and I could not get figures for qwen3.5 and others, so I do not use them to compare generations. Download counts also cannot distinguish who is downloading (individuals, companies, or automatic fetches from other software).
Was it talked about? Events and reactions
| Date | Event | Hacker News score |
|---|---|---|
| 2026-02-03 | Qwen3-Coder-Next (post date) | 735 |
| 2026-02-16 | Qwen3.5 (post date) | 434 |
| 2026-03-03 | Head steps down (post date) | 14 |
| 2026-04-16 | Qwen3.6-35B-A3B (post date) | 1,274 |
| 2026-04-22 | Qwen3.6-27B (post date) | 993 |
| 2026-08-03 | Qwen3.8-Max general availability (post date) | 1,124 |
| 2026-08-12 | Qwen3.8-2.4T (post date) | 713 |
| 2026-08-14 | Qwen3.8-27B (post date) | 1,438 |
| 2026-08-26 | Qwen3.8-Flash-Next (post date) | 704 |
| 2026-09-08 | Qwen3.8 27B quantization benchmark (post date) | 286 |
The Hacker News score is a rough measure of how much overseas developers paid attention. In 2026, several posts beat Qwen3 from Part 2 (869 points). The highest is Qwen3.8-27B at 1,438 points.
Connection to this blog’s measurements
This blog has measured Qwen3.8-27B, the small Qwen3.5 models, and Qwen3.8-Flash-Next on different days and on different machines. Because the conditions differ, I split them into three tables.
| Model | Format | Size [GiB] | Reading speed [tok/s] | Writing speed [tok/s] |
|---|---|---|---|---|
| Qwen3.8-27B | UD-Q4_K_M | 15.32 | 355.79 | 13.02 |
The conditions are EVO-X2 (built-in Radeon 8060S, Vulkan), llama.cpp b11192, median of 5 runs (September 28, 2026). The MoE model Nex-N2.5-mini (34.7B), measured in the same table, wrote at 59.02 tok/s, about 4.5 times the speed of the 27B. 13 tok/s means writing about 13 tokens per second; as a rough guide, Japanese text comes out at roughly a dozen to a couple of dozen characters per second.
| Model (Q4_K_M) | Size [GiB] | Coding tasks [/7] | Writing speed [tok/s] (EVO-X2 built-in GPU) |
|---|---|---|---|
| Qwen3.5-4B | 2.54 | 7 | 65.4 |
| Qwen3.5-2B | 1.18 | 4 | 135.8 |
| Qwen3.5-0.8B | 0.49 | 3 | 245.9 |
The conditions: speed is EVO-X2, llama.cpp b11192, median of 5 runs (September 29, 2026); the coding tasks were run on an RTX 3090 with Ollama, temperature 0, once each. Qwen3-4B, the previous version of the same line (the table in Part 2), scored 6/7 at 77.5 tok/s, so Qwen3.5-4B passed all 7 tasks at 4B. Its writing speed is slower than Qwen3-4B (65.4 vs 77.5 tok/s), and the size is slightly larger.
| Model | Machine / GPU | Writing speed [tok/s] |
|---|---|---|
| Qwen3.8-Flash-Next (2-bit version) | RTX 5060 Ti 16GB (run with Strata; middle 3 of 5 runs) | 70.7 |
The Qwen3.8-Flash-Next value comes from the measurement in this blog’s Qwen3.8-Flash-Next article.
As for comparing the RTX 3090 and EVO-X2, there is an earlier record: on dense mid-sized models the RTX 3090 was about 3 times faster (qwen3.6:27b at 38.6 vs 12.1 tok/s), while on huge MoE models the unified-memory EVO-X2 was faster. I have not re-checked the measurement conditions of this record for this article.
Discussion: how has Qwen changed?
From the published facts, I can say three things.
- In 2026, Qwen split its release terms between the flagship and the sizes that run at home. The Qwen3.7 flagship weights were not released, and with Qwen3.8 the flagship 2.4T was released under a custom license and the 27B under Apache-2.0.
- The releases of the 27B and 35B-A3B are at the center of what people use. Qwen3.8-27B stands out in both downloads and likes.
- The architecture became a hybrid with swapped-in attention, and Flash-Next released the next generation’s design ahead of time.
These are a sequence of facts. I have not confirmed any official explanation of the company’s policy or intent, so I have not written guesses.
Limits: what I could not confirm
- Because I could not read Qwen’s official blog (qwen.ai), I have not confirmed the official release dates of Qwen3.5, 3.6, 3.7, and 3.8-Max, or the details of Qwen3.7.
- The 2.4T custom license was checked through a summarization tool, and I have not finished comparing it with the original text.
- I could not find an official explanation of why the 27B was released.
- I could not get some Ollama pull counts, such as qwen3.5, or third-party evaluations in Japanese.
- This blog’s measurements differ in machine and measurement date for each model, so they are not a comparison under the same conditions.
Summary: what Part 3 showed, and the end of the series
After Qwen3.7 stopped releasing the flagship, Qwen in 2026 released the 2.4T flagship under a custom license and the 27B under Apache-2.0 in August 2026. Qwen3.8-27B was downloaded about 7.12 million times in the last 30 days and has 17,637 likes, a response that stands out even among the Qwen models of 2026.
As a next step, if you want to try it on your own PC, install `qwen3.8:27b` (about 18GB) with Ollama. It is a size that fits if you have a 24GB GPU. Looking across the series, Qwen went from a 7B in 2023 to a lineup running from small to huge models, and in 2026 has reached the point where the 27B that runs at home is at the center of use.
References
- QwenLM/Qwen3.8-Flash-Next (GitHub)
- Qwen3.8-27B (Hugging Face)
- Qwen3.8-2.4T-A95B (Hugging Face)
- qwen3.8 (Ollama library)
- Alibaba’s Qwen team releases Qwen 3.8 models (The Decoder. The headline says both models are Apache-2.0, but per Hugging Face the 2.4T has a custom license)
- Qwen3.8-27B (Hacker News)
- Qwen (Wikipedia, secondary source)