The Most-Downloaded AI Models in the World Are Alibaba's

SCIENCE & TECHNOLOGY · SEPTEMBER 21, 2026

On 18 September 2026, the Qwen team at Alibaba Cloud released Qwen3.8-Omni-Flash, a model that accepts text, images, audio and video through one endpoint, carries a one-million-token context window, and — according to Alibaba's own figures — scores more than 25 percent better on average than the model it replaces, across 29 evaluations. It went live the same day on Qwen Chat, Alibaba Cloud's Model Studio and an OpenAI-compatible API, alongside two pieces of companion tooling, Qwen-MM-Plugins and Qwen-Live Harness, for long-running and real-time work.

I had not been following Alibaba's AI program closely, which turns out to have been a mistake of some size: at the time of this release, Qwen is the most-downloaded family of open-weight AI models in the world, by a margin that is not close. So this entry is two things at once — what is actually new in this specific model, and what the program behind it looks like for anyone who, like me, was reading about three other labs instead.

What "Omnimodal" Actually Means, as Opposed to Multimodal

Most models people call multimodal are a language model with sense organs bolted on, and often the bolting is a pipeline: speech goes to a transcription model, the transcript goes to the language model, the language model's reply goes to a text-to-speech model. Each hop is a lossy translation into text. Tone, overlapping speakers, the pause before someone answers, which of two people on screen was talking — all of that is gone by the time the reasoning starts, because text is the only thing that survives the handoff.

A native omnimodal model is trained so that audio and video are first-class inputs in the same representation the text lives in, rather than guests who have to be translated at the door. Qwen has been building toward this for a while: Qwen2.5-Omni introduced a "Thinker-Talker" design — a Thinker that perceives text, images, audio and video and produces reasoning and text, and a Talker that turns the Thinker's stream into speech tokens in real time — and Qwen3-Omni carried it forward as an openly published model. So the lineage here is genuine architecture work, not a marketing word applied retroactively.

Worth one honest asterisk, though: this particular model's native output is text. It sees and hears everything; it answers in writing. There is a separate -Realtime variant aimed at sub-1.4-second streaming interaction, but the headline model is four modalities in, one modality out. If the phrase "omnimodal" made you picture a machine that talks back in its own voice, that is the earlier Omni line's trick, not this one's job.

The 26 Percent, and What It Is Actually Measuring

First, the number itself needs a small correction, and it is the kind of thing worth getting right rather than repeating: Qwen's own announcement states more than 25 percent average improvement across 29 evaluations versus its predecessor Qwen3.5-Omni-Plus. Some of the coverage rounds this to "more than 26 percent across 30 evaluations." The difference does not change anything material, but a self-reported average across an evaluation suite the vendor chose is a soft number twice over, and it is more useful to look at what moved inside it.

The individual results are lopsided in a way the average hides. The biggest single jump is on WildClawBench-MM, a multimodal agent benchmark, where the model scores 71.0 against its predecessor's 34.5 — a 36.5-point gain, which is less "improved" than "the previous model was not really doing this task." AgenticVBench moves 22.3 points. Speaker diarization — the problem of working out who said which sentence in a recording with several people in it — goes from a roughly 88 percent error rate on the AliMeeting set to about 3 percent, which again reads as a capability appearing rather than sharpening.

Against Google's Gemini 3.8 Flash, the comparison Alibaba chose to publish, it is a split decision: Qwen ahead on the diarization sets, on music reasoning, on the joint audio-video benchmark and on multimodal agent work; Gemini ahead on Video-MME-v2, 71.0 to 65.0. A vendor that publishes the benchmark it loses is at least showing you the whole table.

Why a Million Tokens Matters Specifically Here

Long context windows have become a slightly tired bullet point on text models, where the honest question is whether anyone actually has a million tokens of text they want answered about. Video is different. Video is enormous. An hour of footage sampled at a usable frame rate, plus its audio track, plus a transcript, eats context at a rate a chat log never approaches — which is exactly why the published limits read the way they do: video up to two hours or two gigabytes, audio up to three hours, stable at 15 frames per second, with the million-token budget split roughly 991,000 tokens of input against 131,000 of output.

Put those together and the shape of the intended product is obvious. This is not a model for answering questions about a paragraph. It is a model for handing over a recorded meeting, a lecture, a day of security footage or a long interview and getting back something that required watching the whole thing. The 98-percent cut in the per-hour price of audio input (93 percent for audio plus video) is the same argument made in the currency that decides whether anyone does it twice.

Why the Company You Know for E-Commerce Is a Serious Lab

The short version: Alibaba is not primarily an AI lab that happens to sell things, it is a cloud provider that needs a reason for you to rent its servers, and a very good free model is the most effective advertisement a cloud provider has ever found.

The program started in April 2023 as Tongyi Qianwen and went public that September after Chinese regulatory approval. Qwen2 landed in June 2024, Qwen2.5 that September, Qwen3 in April 2025 — trained on 36 trillion tokens across 119 languages — and the Qwen3.8 generation arrived in August 2026, topped by Qwen3.8-Max at 2.4 trillion parameters, which at the time made it the second-largest open-weight language model in the world, behind Moonshot AI's Kimi K3.

The distribution numbers are what actually surprised me. Hugging Face's state-of-open-models report in mid-August 2026 put Qwen past three billion downloads in six months, against roughly 418 million for Google's open models and 227 million for Meta's over 2026 — more than both combined, several times over. Alibaba has open-sourced over 460 models; the ecosystem has produced more than 300,000 derivatives. Chinese open-weight models, led by Qwen and DeepSeek, now account for something like 30 percent of global model downloads. Meanwhile Alibaba's Cloud Intelligence Group pushed capital expenditure to ¥126.1 billion for fiscal 2026, with AI running at roughly 30 percent of external cloud revenue, and in July 2026 Qwen became the model behind Apple Intelligence in China, where ChatGPT is not available.

That is the strategy in one sentence: give away the weights, win the developers, rent them the servers. It is the same play Google ran with Android, and it is being run against American frontier labs whose best models are not downloadable at all.

Where Qwen Sits Against DeepSeek and the Rest

Among the Chinese open-weight labs — DeepSeek, Moonshot's Kimi, MiniMax, Zhipu — the rough division is that DeepSeek competes on licence permissiveness and cost per token and has the louder reputation for raw reasoning, while Qwen competes on breadth: more sizes, more modalities, more languages, and an ecosystem large enough that fine-tuning a Qwen checkpoint is the path of least resistance for most people building something. Against the Western labs, the comparison is harder to make honestly, because the Western frontier models are closed and the benchmark tables everyone quotes are published by whoever is being compared.

Which is the caveat to carry out of all of this. Downloads measure distribution, not capability — a model that is free, small enough to run locally and permissively licensed will out-download a better model that is none of those things, every time. And a 26 percent average lift across a vendor's own evaluation suite tells you the vendor believes its model improved; it does not tell you by how much on your work.

The Part That Is Not in the Announcement

Three things worth knowing before anyone builds on this. The weights are not open. Qwen3.8-Omni-Flash is API-only, running on Alibaba's own infrastructure, even though it is built on Qwen3.8-Flash-Next, whose weights were published in August 2026. The open-weight strategy has a boundary, and the newest omnimodal model sits on the far side of it. Video and audio limits are real limits — two hours, two gigabytes, three hours of audio, 15 frames per second — and anything past those needs chunking you have to build yourself. And early reports of the companion tooling are mixed: the Qwen-MM-Plugins integration with third-party coding agents has been described as prone to interruptions and unstable sessions, which is the sort of thing that gets fixed in a fortnight but is worth knowing on day three rather than discovering on day one.

Where I Could Be Wrong

• Every benchmark figure here is the vendor's. Alibaba selected the evaluation suite, ran the comparisons, and chose which competitor results to print. Nothing in this entry is an independent replication, and the history of this industry is that self-reported margins shrink when someone else runs the same tests on their own workloads.
• The headline improvement figure is reported two ways. Qwen's own announcement says more than 25 percent across 29 evaluations; several outlets say more than 26 percent across 30. I have used the company's own smaller number, but I have not resolved which suite the larger figure counts, and it is possible the two refer to different cuts of the same results.
• The language-coverage figures disagree between sources — one summary gives 74 languages plus 39 Chinese dialects for speech recognition and a smaller set for generation, another states 113 languages outright. I have therefore avoided quoting a single language count in the body, because the two are not reconcilable from the public material.
• I have not used this model. Everything above is read, not tested. The gap between a benchmark table and a model's behaviour on a real two-hour recording is precisely the gap that benchmarks exist to hide.
• "Native omnimodal" is a claim about architecture I am taking on trust. The published material describes the design at the level of capability rather than internals, and I have leaned on the documented Thinker-Talker lineage of the earlier Omni models to describe what the family means by the term. A model can be described as native and still be doing more internal translation than the word implies.
• Download counts are a soft metric in a second way I did not dwell on: automated pipelines, CI systems and mirrors pull models repeatedly, and no public report I have seen separates a human evaluating a model from a build server fetching it for the hundredth time.

Sources

1. Qwen (Alibaba Cloud). Qwen3.8-Omni-Flash: Omni Senses. Agentic Delivery. 18 September 2026. qwen.ai
2. TechNode. Alibaba's Qwen releases Qwen3.8-Omni-Flash with 1M-token context. 18 September 2026. technode.com
3. MarkTechPost. Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understanding and Tool Use. 18 September 2026. marktechpost.com
4. DataCamp. Qwen3.8-Omni-Flash: Features and Benchmarks. September 2026 — source of the benchmark-by-benchmark comparison against Gemini 3.8 Flash. datacamp.com
5. Fortune. Alibaba AI models hit 3 billion downloads, passing Meta, Google. 15 August 2026, reporting Hugging Face's state-of-open-models figures. fortune.com
6. Qwen Team, Alibaba Group. Qwen2.5-Omni Technical Report. arXiv:2503.20215 — the Thinker-Talker architecture. arxiv.org
7. Qwen Team, Alibaba Group. Qwen3-Omni Technical Report. arXiv:2509.17765. arxiv.org
8. Wikipedia. Qwen. Accessed September 2026 — release timeline and model-generation dates. en.wikipedia.org