Qwen 3.8 in one line: Alibaba announced a 2.4 trillion-parameter multimodal model on 19 July 2026, called it second only to Claude Fable 5, and shipped a paid preview the same day. What it did not ship was a benchmark table, a model card, a license, a context window figure, or any weights. If you came here to self-host it, you cannot - and this post explains what to run instead.
Here is the honest state of it.
Quick facts
| Item | Status |
|---|---|
| Announced | 19 July 2026, WAIC Shanghai |
| What is live | Qwen3.8-Max-Preview (proprietary, paid) |
| Total parameters | 2.4 trillion (Alibaba's own figure) |
| Active parameters per token | Not disclosed |
| Architecture | Not disclosed |
| Context window | Not disclosed |
| Published benchmarks | None |
| Open weights | Promised "soon", no date, no license |
| Modalities | Text, images, video, documents (claimed) |
Every performance claim below traces to a single post on X. No independent evaluation exists yet.
What Alibaba actually claimed
The announcement came from the official Qwen account on 19 July:
"Qwen3.8 is launching and going open-weight soon! With a massive 2.4T parameters, this model is continuously evolving. We believe it's one of the most powerful model available today, compatible to leading frontier AI models, second only to Fable 5."
That is the entire basis for the "second only to Claude Fable 5" headline you have seen everywhere. It is a marketing line in a tweet, not a result. Note that four separate outlets ran that claim, and all four cite the same post. For context on the model it is measuring itself against, see our Claude profile.
Qwen developer Shuai Bai added that the model "should beat Qwen3.7-Max on coding, full-stack development, data analysis, and office workflows", and described it as Alibaba's first multimodal model above one trillion parameters.
That is the sum total of the performance information available.
The coding angle: what is claimed vs what is measurable
If you write code for a living, this is the section that matters, so let me be blunt about it.
There is not a single coding benchmark number for Qwen 3.8. No SWE-bench. No LiveCodeBench. No Aider polyglot. No HumanEval. The claim is directional - better than the previous Max model - with nothing behind it.
Compare that to how a serious coding-model release usually looks. When a lab wants developers to switch, it publishes SWE-bench Verified alongside the pricing, because that is the number engineering teams actually budget against. Alibaba published pricing and skipped the number.
What you can do today is test it yourself. The preview is reachable through Alibaba's Token Plan subscription and through Qoder and QoderWork, Alibaba's agentic coding platforms. Alibaba says the preview runs at roughly 10% of the standard credit rate during the day and less overnight, with Token Plans starting around RMB 35 per month.
That cheap preview rate is the genuinely interesting part of this launch. It is inexpensive enough to run your own evaluation on your own repository, which is worth more than any benchmark table someone else publishes. Point it at a real ticket from your backlog and compare the diff to whatever you use now.
What you should not do is move production traffic based on a tweet.
The self-hosting angle: why 2.4T is not a home lab model
Here is the part the launch coverage skipped entirely.
Even when the weights land, running a 2.4 trillion-parameter model yourself is out of reach for nearly everyone. Do the arithmetic. At 4-bit quantization, 2.4 trillion parameters need roughly 1.2 TB of memory for the weights alone, before you account for the KV cache, activations, or any serving overhead.
An NVIDIA H200 has 141 GB. You would need around nine of them, networked, just to hold the weights - and that is the optimistic 4-bit case.
There is also a detail nobody has disclosed that changes this maths completely: the active parameter count. Qwen's recent models are sparse mixture-of-experts designs, where only a fraction of parameters fire per token. Qwen3-235B-A22B activates 22 billion of its 235 billion. If Qwen 3.8 follows that pattern, the compute per token could be far more modest than 2.4T suggests - but you still need all 2.4T resident in memory. Alibaba has not said which experts activate, how many, or how the model is structured at all.
So "2.4 trillion parameters, going open-weight" reads as a headline about accessibility. In practice it describes a model that only well-funded infrastructure teams will ever run.
What to self-host instead
If open weights are what you actually need, these exist today and you can download them now:
- Qwen3.5 (397B) - Alibaba's own open-weight release from February 2026. Large but genuinely deployable on a serious multi-GPU setup, and it is from the same family, so prompt patterns carry over.
- Kimi K3 (2.8T) - Moonshot AI released this on 17 July 2026, two days before the Qwen 3.8 preview, with open weights and a real technical write-up. See the Kimi profile, and more on this below.
- DeepSeek - not the largest, but the most battle-tested open-weight line for reasoning and code, and cheap to run. Details on the DeepSeek profile.
- Smaller Qwen3 variants - the 30B-A3B class activates around 3B parameters per token and runs on hardware that a small team can actually afford.
For a broader view of the free and open end of the market, see our Qwen Chat tool profile and our roundup of free AI tools for small business.
The Kimi K3 comparison nobody is making
This is the most useful context for judging the Qwen 3.8 announcement, and it went almost unreported.
Two days earlier, Moonshot AI released Kimi K3: 2.8 trillion parameters, larger than Qwen 3.8, and actually open-weight rather than promised. More importantly, Moonshot published the technical detail Alibaba withheld.
| Qwen 3.8 | Kimi K3 | |
|---|---|---|
| Parameters | 2.4T (claimed) | 2.8T |
| Weights available | No | Yes |
| Architecture published | No | Yes - Kimi Delta Attention, Attention Residuals |
| Expert routing | Not disclosed | 16 active of 896 |
| Context window | Not disclosed | 1M tokens |
| Third-party leaderboard result | None | Topped Arena's Frontend Code leaderboard |
| Pricing model | Credit subscription | Flat pay-as-you-go |
Kimi K3 is the first open-weight model to top Arena's Frontend Code leaderboard - a third-party result, not a self-report. If your interest in Qwen 3.8 was driven by the open-weight promise or by coding performance, K3 is the model that has already delivered both.
One caveat: Moonshot reportedly paused new signups shortly after launch because of GPU capacity. Availability may be uneven.
Should you switch?
No - not yet, and not on this evidence.
Test it if you are already paying for Alibaba's ecosystem, you use Qoder, or the 10% preview rate makes a real evaluation cheap for you. Run it against your own tasks and judge the output.
Wait if you need reproducible benchmarks, a license you can read, a known context window, or weights you can deploy. None of those exist right now.
Look at Kimi K3 instead if open weights were the whole appeal.
There is also a pattern worth knowing. Analysts covering this launch pointed out that previous Qwen releases followed the same shape - strong comparative claims against frontier models, published without third-party verification. That does not make Qwen 3.8 bad. Qwen models are genuinely good and the family is one of the strongest open lines in the world. It means the claim and the evidence arrived at different times, and only the claim has shown up so far.
The practical rule holds regardless of vendor: your own evaluation on your own workload beats any launch-day number.
FAQ
Is Qwen 3.8 open source? Not yet. Alibaba says weights are coming "soon" but has named no date and no license. The preview available today is proprietary and paid. There is no Hugging Face repository for it. Note that this would also reverse Alibaba's recent pattern - Qwen3.7-Max was never open-weighted.
Can I self-host Qwen 3.8? No. The weights have not been released. Even once they are, 2.4 trillion parameters need roughly 1.2 TB of memory at 4-bit for the weights alone - around nine H200 GPUs before overhead. For practical self-hosting today, look at Qwen3.5 (397B) or the smaller Qwen3 variants.
Is Qwen 3.8 free? No. Access runs through Alibaba's paid Token Plan subscription, or through Qoder and QoderWork. Alibaba says the preview is discounted to about 10% of the standard credit rate, with plans starting near RMB 35 per month. The separate consumer Qwen Chat app remains free but does not run this model.
How do I access Qwen 3.8? Through Alibaba's Token Plan - qwencloud.com for international users, platform.qianwenai.com inside China - or through the Qoder and QoderWork coding platforms. Some coverage claimed it was mainland-China only; Alibaba's own announcement lists both endpoints.
Is Qwen 3.8 better than Claude? Unknown. Alibaba says it is "second only to Fable 5", which is a self-assessment with no benchmark behind it and no independent evaluation. Until Artificial Analysis, LMArena, or a comparable third party publishes numbers, treat the comparison as marketing.
How good is Qwen 3.8 at coding? No coding benchmarks have been published. A Qwen developer said it should improve on Qwen3.7-Max for coding and full-stack work. If coding is your use case, the cheap preview rate makes it affordable to test on your own repository, which is the only measurement that matters for your workload.
What is Qwen 3.8's context window? Not disclosed. Alibaba published no context length. By comparison, Kimi K3 documents a 1 million token window.
Facts current as of 20 July 2026. Parameter counts and performance comparisons are Alibaba's own figures and remain independently unverified. We will update this post when benchmarks, a license, or weights are published.