Qwen3.8 is the newest flagship model from Alibaba’s Qwen team, previewed on July 19, 2026 at the World AI Conference in Shanghai. Alibaba describes Qwen3.8 as a 2.4-trillion-parameter multimodal model and says it ranks second only to Anthropic’s Claude Fable 5 among the systems it tested. Both of those figures are the company’s own claims. What shipped on launch day was a paid preview, not a benchmark table, a model card, or a license. This post separates what Alibaba has confirmed about Qwen3.8 from what it has so far only asserted, and places the model among China’s fast-growing roster of open-weight releases.
What Alibaba actually announced
The Qwen account posted that Qwen3.8 was launching and would go open-weight "soon," calling it one of the most powerful models available today and comparable to leading frontier systems. The build that went live is a preview, named Qwen3.8-Max-Preview, and it is real and purchasable rather than a slide in a keynote.
Qwen developer Shuai Bai added the technical framing. He described Qwen3.8 as the team’s first multimodal model above 1 trillion parameters, one that processes text, images, video, and documents together. Alibaba says the model should beat its predecessor, Qwen3.7-Max, on coding, full-stack development, data analysis, and office workflows. The preview is also compatible with the OpenAI and Anthropic API protocols, so existing coding agents can point at it without rebuilding their harnesses.
Three things are confirmed: the preview exists and is for sale, it is a sparse mixture-of-experts multimodal system, and it speaks the common API protocols. Almost everything else that made headlines is a claim, and the sections below treat it that way.
The size question, and the number nobody has published
The 2.4-trillion-parameter figure is Alibaba’s own. No model card or specification confirms it, and total parameter count is not the same as usable compute. That distinction matters more than the headline number, and Qwen’s own history proves the point.
Qwen’s earlier mixture-of-experts models make the pattern clear. Qwen3-235B-A22B carries 235 billion total parameters but activates only 22 billion per token. Qwen3-30B-A3B activates roughly 3 billion. The Max tier uses the same sparse design. For Qwen3.8, the active-parameter count is the single number that decides real serving cost, and Alibaba has not disclosed it. Without that figure, 2.4 trillion total parameters says very little about what the model costs to run.
The raw scale is also awkward for anyone hoping to self-host. A 2.4-trillion-parameter model at 4-bit precision needs roughly 1.2 terabytes just for the weights, before the memory overhead of serving. A single Nvidia H200 accelerator carries 141 gigabytes, so even eight of them leave uncomfortable math. The practical question is therefore not leaderboard position but whether Alibaba ships a smaller activated variant, a well-made quantized checkpoint, or a distilled sibling that a workstation can actually load. On context length, the verified baseline is the predecessor: Qwen3.7-Max handled a 1 million token window, and Qwen3.8 is expected to match or extend that, though the official spec is not out yet.
"Open weights" is a promise, not a release
Alibaba said Qwen3.8 would go open-weight "soon." That is the entire commitment. There is no date, no license text, and no Hugging Face or ModelScope repository for the model as of this writing. Anyone naming a specific week is guessing.
This would also break a pattern. Qwen’s smaller dense models have long shipped under the permissive Apache 2.0 license on Hugging Face, ModelScope, and Kaggle, which is a large part of why the Qwen family became a favorite among open-source builders. The Max-tier flagships, by contrast, have historically shipped closed. So an open-weight Qwen3.8 would be a genuine shift, not a continuation, which is one reason to wait for the repository before treating the promise as done.
It is also worth being precise about the language. "Open weights" means the trained model file can be downloaded and run. It does not automatically mean open source in the strict sense: the training data and code are usually not released, and the accompanying license is often a custom one with usage restrictions rather than an approved open-source license. When the Qwen3.8 weights do land, the license file will decide what you are actually allowed to do with them, especially for commercial use. If you want the underlying vocabulary here, our explainer on what a frontier model is covers the open-versus-closed distinction in more depth.
The benchmark claim, and why to treat it as provisional
"Second only to Fable 5" is a striking line, and it rests entirely on Alibaba’s internal evaluation. No benchmark table has been published, no model card accompanies the claim, and no independent lab has verified the ranking. Every capability number circulating for Qwen3.8 is, strictly speaking, still a Qwen3.7-Max number until Alibaba releases real results.
The contrast with a recent rival is instructive. Kimi K3, from Beijing lab Moonshot AI, shipped only days earlier with third-party scrutiny attached: the independent Artificial Analysis index placed it third overall. Qwen3.8 arrived with a superlative and a preview link. That does not make the claim false, but it does make it unverified, and a launch-day superlative from any vendor deserves the same skepticism regardless of where the lab is based.
Before treating Qwen3.8 as a frontier-class option, five concrete things need to land, and each is checkable in a minute once it exists: an official blog post with a benchmark table, the active-parameter count, a Hugging Face repository with a real license file, published API pricing, and independent evaluation from an outlet such as Artificial Analysis or LMArena. Until then, the honest description is a promising preview with an unproven ranking.
Pricing and how to try it today
Qwen3.8-Max-Preview is available now through Alibaba’s Token Plan subscription and its Qoder and QoderWork developer platforms, offered at 10% of standard pricing during the preview. Alibaba has not published a standalone rate card for the model, so the "10% of standard" figure is relative to a standard price that is itself not yet fixed for this model.
For a reference point, the predecessor is useful. Alibaba lists Qwen3.7-Max at roughly $1.25 per million input tokens and $3.75 per million output tokens, and the company’s historical edge has been price-to-performance rather than topping a leaderboard. If Qwen3.8 lands near those numbers while approaching frontier quality, the value case writes itself, but that is a conditional built on unpublished pricing. If you are new to how those two rates differ and why output usually costs more, our primer on input versus output tokens explains the mechanics. The zero-per-token floor, self-hosting, remains unavailable until the weights actually ship.
Where Qwen3.8 fits among China’s open models
The most important context for Qwen3.8 is the company it keeps. Chinese labs have pushed a wave of large mixture-of-experts models into the open-weight tier this year, and Qwen3.8 lands in the middle of that cluster rather than alone at the front. DeepSeek V4 Pro and Zhipu’s GLM-5.2 both compete on the same axis of strong capability at aggressive prices, with weights that developers can inspect and, in most cases, run.
The timing is the sharpest part of the story. Kimi K3, a 2.8-trillion-parameter open-weight model, arrived just two to three days before the Qwen3.8 preview, and Alibaba holds a roughly 36% stake in Moonshot, the lab behind it. Reading Qwen3.8 as a direct answer to Kimi K3 is reasonable. Together these releases push Chinese developers firmly into the multi-trillion-parameter range, a scale until recently associated mainly with the largest US labs.
Against those US frontier systems, the comparison is still lopsided on evidence. Models like Claude Fable 5 and GPT-5.6 ship with published benchmarks and broad third-party testing; Qwen3.8 so far ships with neither. The genuine advantages of the Chinese open cohort are price and openness, not verified peak capability. There is also a nuance that matters for many buyers: the API is China-hosted, which raises data-residency and compliance questions for some organizations, and the open-weight route that sidesteps that concern depends on weights and a license that have not yet been released. For now, Qwen3.8 is best understood as a strong signal of where the open-model race is heading, not a finished product to migrate onto.
Frequently Asked Questions
Is Qwen3.8 open source?
Not yet, and possibly not in the strict sense even later. Alibaba has promised open weights “soon” with no date and no license. Even when the weights arrive, “open weights” means the model file is downloadable, not that the training data, code, and an approved open-source license come with it. The license text will decide what commercial use is permitted.
How many parameters does Qwen3.8 have?
Alibaba claims 2.4 trillion total parameters, but no model card confirms it. More importantly, it is a sparse mixture-of-experts model, so only a fraction of those parameters activate per token. The active-parameter count, which determines real serving cost, has not been disclosed.
Can I download Qwen3.8 from Hugging Face or ModelScope?
Not as of this writing. Smaller Qwen models have long been available on Hugging Face and ModelScope under Apache 2.0, but Qwen3.8 has no public repository yet. Watch for an official Hugging Face page with a real license file before assuming it is downloadable.
Is Qwen3.8 really second only to Claude Fable 5?
That is Alibaba’s own claim, based on internal evaluation. No benchmark table has been published and no independent lab has verified the ranking. Treat it as provisional until outlets like Artificial Analysis or LMArena test it.
How much does Qwen3.8 cost?
The preview is available through Alibaba’s Token Plan, Qoder, and QoderWork at 10% of standard pricing, but Alibaba has not published a full rate card for the model. For reference, the predecessor Qwen3.7-Max is listed at about $1.25 per million input tokens and $3.75 per million output tokens.
What is the difference between Qwen3.8 and Qwen3.8-Max?
Qwen3.8 is the model family, and Max is its top flagship tier. The build shipped on July 19 is specifically Qwen3.8-Max-Preview, a preview release of that flagship. Any smaller or distilled variants that a workstation could run have not been announced.
How does Qwen3.8 compare to Kimi K3 and DeepSeek?
All three are large Chinese mixture-of-experts models competing on price and openness. Kimi K3 and DeepSeek V4 Pro shipped with published details and, in Kimi’s case, third-party benchmarks. Qwen3.8 so far ships as a preview with vendor claims only, so a direct comparison waits on real numbers.
Should I move production workloads to Qwen3.8?
Not on the strength of a preview. Test it on your own workload through the official console, keep production traffic where it is, and wait for a benchmark table, the active-parameter count, a license, and independent evaluation before migrating.
