Qwen Commits to Keeping Future Models Open-Weight
Alibaba's Qwen team says Qwen 4, Qwen 5 and future releases will stay open-weight, calming fears sparked by social media rumors.
What did Qwen actually commit to?
The team behind Alibaba’s Qwen model family has publicly stated that upcoming releases, including Qwen 4 and Qwen 5, will continue to ship as open-weight models. The statement, attributed to the head of the Qwen team, pushed back on rumors circulating on social media that Qwen 4 might be the last open release, and framed the decision around a broader principle: that the value of AI shouldn’t sit with a small number of companies. For a community that had spent weeks worrying about where Qwen was headed, it was a direct answer to a direct fear.
TL;DR
- Alibaba’s Qwen team publicly confirmed that Qwen 4, Qwen 5, and future models will remain open-weight, directly addressing rumors that Qwen 4 could be the last such release.
- The announcement matters because access to AI is not evenly distributed. A $20 monthly subscription is a real financial decision in much of the world, and rate limits on hosted tools make things worse.
- Even open-weight models aren’t automatically accessible. Large models still need serious GPU hardware, which puts them out of reach for many individuals and small teams regardless of licensing.
- Smaller and mid-sized Qwen models, like the Qwen3 VL 30B instruct-thinking model, matter because they run on a single consumer GPU or even a laptop once quantized, without giving up much capability.
- That same 30-billion-parameter dense model handles text, images, and video, supports a context window over 200,000 tokens, ships under Apache 2.0, and has reportedly outperformed Claude Opus 4.5 on at least one coding benchmark built from real GitHub issues.
- Qwen’s mixture-of-experts models activate only a fraction of total parameters per token, which keeps compute costs down relative to the full model size.
- Community numbers cited around the announcement include over 3 billion downloads across ModelScope and Hugging Face combined, more than 460 open models released, and over 300,000 derivative models built on top of them.
Nothing gets lost.
Why does open-weight access matter this much?
Most coverage of frontier AI focuses on benchmark scores and pricing tiers for people who can already afford to participate. That leaves out a large chunk of the world. Outside wealthier countries, a monthly subscription to a hosted AI product is a genuine budget decision, not a rounding error. Rate limits and usage caps compound the problem: someone finally gets a workflow built around a tool, then gets locked out mid-task because they hit a quota.
Open-weight models are often presented as the fix for this, and sometimes they are. But a model being “open” doesn’t automatically make it usable. A massive open-weight model still needs GPUs to run, and those GPUs can cost more than many people earn annually. Renting compute from cloud providers helps, but it’s still a real and recurring expense for people in lower-income regions. In that sense, a huge open-weight model can be open in name only if nobody outside a well-funded lab can actually run it.
What makes Qwen’s smaller models different?
This is where Qwen’s strategy stands out. Rather than only chasing frontier-scale parameter counts, the team has kept releasing small and mid-sized models that run on modest hardware, sometimes a single consumer GPU, sometimes a laptop, without a steep drop in quality.
A good example is Qwen3 VL 30B instruct-thinking, a dense 30-billion-parameter model released a few months before the open-source commitment. It handles text, images, and video, supports a context window beyond 200,000 tokens, and is released under the Apache 2.0 license, which permits commercial use. That license detail matters: developers can download the model, run it privately, and build a business on top of it without licensing friction.
Once quantized to 4-bit precision, the model reportedly fits on a single graphics card, which is why it has become a common choice for people running local inference through tools like LM Studio or llama.cpp. On SWE-bench Pro, a coding benchmark built from real-world GitHub issues, it has reportedly outperformed Claude Opus 4.5 on one measure while trailing on another. Whatever the exact numbers, the broader point holds: a 30-billion-parameter model showing up at all in comparisons against a major closed frontier model says something about how far efficient open models have come.
How does the mixture-of-experts approach fit in?
Alongside dense models like the 30B instruct-thinking release, Qwen has also published a range of mixture-of-experts (MoE) models. In an MoE architecture, only a subset of the model’s total parameters activates for any given input token. That means the effective compute cost per token is much lower than the total parameter count would suggest, letting a model with a large overall size run more cheaply than a dense model of similar scale.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Across its full lineup, Qwen has released several hundred models to date, spanning dense and MoE architectures at multiple sizes. That range, from sub-billion-parameter models up to large MoE systems, is part of why the family has been adopted so widely: developers can pick a size that matches their hardware instead of being forced into a one-size-fits-all release.
How big is the Qwen community, really?
The scale of adoption is one of the stronger arguments for why this open-source commitment matters. Qwen models have reportedly been downloaded more than 3 billion times combined across ModelScope and Hugging Face. The team has released over 460 open models to date, and the community has built more than 300,000 derivative models on top of them, fine-tunes, merges, and adaptations published by other developers and researchers.
Those derivative models represent real usage: students experimenting with fine-tuning, startups building products on a base they didn’t have to train from scratch, and independent developers testing ideas that would be financially out of reach if every experiment required renting frontier-scale compute from a closed API provider.
Is this commitment something to fully trust yet?
It’s worth being clear-eyed here. A public statement is not the same as a multi-year track record, and the Qwen team itself has had rocky moments. At one point, part of the original Qwen team reportedly left the project after the group scaled back releases of smaller models (sizes like 0.5B, 4B, and 9B parameters), which had been popular with developers running on limited hardware. That history is a reasonable reason for measured optimism rather than certainty.
It’s also worth separating “open-weight” from “open-source” as terms. Open-weight generally means the trained model parameters are published for anyone to download and run, often under a permissive license like Apache 2.0. Open-source, in the stricter software sense, usually implies the training code, data, and methodology are available too, which is rarely the case even for “open” AI models. Qwen’s commitment, as described, is about continuing to release open weights, not necessarily full training transparency. That distinction matters for anyone citing this as a broader “open-source AI” milestone.
Still, a public commitment backed by years of consistent releases, hundreds of models, and a track record of supporting small and mid-sized architectures is meaningfully different from a vague promise. Qwen isn’t the only team releasing open-weight models under permissive licenses (there are notable efforts from the US, China, and Europe), but the scale and consistency of its releases is part of why this particular announcement landed as reassuring news rather than routine PR.
Frequently Asked Questions
What exactly did the Qwen team promise?
The head of the Qwen team stated publicly that Qwen 4, Qwen 5, and future models will continue to be released as open-weight, directly countering rumors that Qwen 4 might be the last open release in the lineup.
Is “open-weight” the same as “open-source”?
Not exactly. Open-weight means the trained model parameters are published for download and use, often under licenses like Apache 2.0. Open-source typically implies broader transparency, including training code and data, which most open-weight AI releases, including Qwen’s, don’t fully provide.
Why do small and medium Qwen models matter more than huge ones for accessibility?
Without shadow IT.
Large open-weight models still require expensive GPU hardware to run, putting them out of reach for many individuals and small teams even though the weights are freely available. Smaller and mid-sized Qwen models, including ones that fit on a single consumer GPU once quantized, let far more people actually use the technology.
What is the Qwen3 VL 30B instruct-thinking model?
It’s a dense 30-billion-parameter multimodal model from Qwen that handles text, images, and video, supports a context window over 200,000 tokens, and is released under the Apache 2.0 license, allowing commercial use. Quantized to 4-bit, it reportedly runs on a single graphics card.
How many models has Qwen released so far?
According to figures cited around this announcement, Qwen has released more than 460 open models to date, with combined downloads exceeding 3 billion across ModelScope and Hugging Face, and over 300,000 community-built derivative models.
