⏱️ Lectura: 9 min
Alibaba raised the stakes in the open model race: the official Qwen account announced today, July 19, the launch of Qwen3.8, a 2.4 trillion parameter model that will soon arrive in open weights. According to the post, the system is “one of the most powerful models available today” and only trails Fable 5, Anthropic’s flagship model, although there are no public benchmarks yet to confirm this.
📑 En este artículo
A preview version, Qwen3.8-Max-Preview, is already active on Alibaba Cloud’s Token Plan and on the Qoder and QoderWork tools, so any developer can try it before the weights are released.
TL;DR
- Alibaba announced Qwen3.8 on July 19, 2026 on X, confirming it will arrive with open weights soon.
- According to the @Alibaba_Qwen account, the model has 2.4 trillion parameters.
- The Qwen3.8-Max-Preview version is already available on the Token Plan, Qoder, and QoderWork.
- Alibaba claims Qwen3.8 only trails Fable 5, from Anthropic, without public benchmarks to back it up.
- Access via Token Plan has separate pricing for the international market (qwencloud.com) and China (platform.qianwenai.com).
- The API is compatible with the OpenAI format through DashScope’s compatible mode.
- The launch adds to a recent wave of open Chinese models like DeepSeek V4, Kimi K3, and GLM-5.2.
Introduction
The Qwen family is Alibaba’s line of language models, released since 2023, with versions ranging from a few billion parameters to frontier-scale models. Qwen3.8 is the latest release and, according to the announcement itself, the largest so far in declared parameters.
The announcement came via X, without a technical paper or a published model card yet. That matters: the numbers circulating today are self-reported by Alibaba, not verified by an independent third party.
What happened with Qwen3.8
At 8:29 AM on July 19, the @Alibaba_Qwen account posted that Qwen3.8 “is launching and going open-weight soon,” with 2.4 trillion parameters and “continuously evolving.” The post adds that the model is “comparable to the leading frontier models, second only to Fable 5.”
Along with the announcement, Alibaba enabled Qwen3.8-Max-Preview, an early version accessible through Alibaba Cloud’s Token Plan and the Qoder and QoderWork development tools. No date was given for the release of the open weights, nor a specific license: previous Qwen releases used Apache 2.0-type licenses or Qwen’s own license depending on the model size.
Context and history
Qwen isn’t a new project. Alibaba open-sourced the weights of its first Qwen models in 2023 and has since released Qwen2, Qwen2.5, and Qwen3 in succession, each with variants ranging from under a billion to hundreds of billions of parameters. The family’s models are typically distributed on Hugging Face under the official Qwen organization, and the reference code lives on GitHub, under QwenLM.
The Qwen3.8 announcement adds to a string of large-scale open Chinese models released in recent months, including DeepSeek V4, Moonshot AI’s Kimi K3, and Zhipu’s GLM-5.2. All of them are competing to position themselves close to the frontier set by closed models from labs like Anthropic, OpenAI, and Google. The difference with Qwen3.8, at least for now, is that Alibaba explicitly presents it as comparable to that frontier, not just as a cheaper alternative.
Technical details and performance of Qwen3.8
The central figure in the announcement is the parameter count: 2.4 trillion. Alibaba didn’t specify whether it’s a dense model or a mixture-of-experts (MoE) architecture, the design used in its previous large releases to keep inference costs manageable. Nor did it publish results on standard benchmarks like MMLU, GPQA, or SWE-bench, or any direct, verifiable comparison against other models.
The phrase “continuously evolving” suggests Qwen3.8 could be updated over time without a name change, similar to what other labs do with preview versions that receive incremental improvements before stabilizing.
📌 Note: The “second only to Fable 5” comparison is a claim made by Alibaba in its own announcement, not the result of an independent benchmark published to date.
For a reference on where Qwen3.8 sits among current access options, here’s how the channels Alibaba enabled alongside the announcement break down:
| Channel | Audience | What it offers | Limitation |
|---|---|---|---|
| International Token Plan | Developers outside China | Access to Qwen3.8-Max-Preview via token consumption through qwencloud.com | Pricing in dollars, requires an international account |
| China Token Plan | Developers in China | Same model via platform.qianwenai.com | Billing and accounts separate from the international plan |
| Qoder | Development teams | Integration of Qwen3.8-Max-Preview into the coding workflow | Early beta, no public SLA |
| QoderWork | Collaborative/enterprise use | Model access within team workflows | Public documentation limited at the time of the announcement |
How to start testing Qwen3.8
As long as there are no open weights to run the model locally, the only way to test Qwen3.8 is against Alibaba Cloud’s API. Qwen models expose an OpenAI-compatible mode through DashScope, which lets you reuse any client that already speaks that protocol.
pip install openaiWith the library installed, all you need is to point the base_url to Alibaba Cloud’s compatible endpoint and use the preview model identifier:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_DASHSCOPE_API_KEY",
base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)
response = client.chat.completions.create(
model="qwen3.8-max-preview",
messages=[
{"role": "user", "content": "Summarize in 3 lines what changes with Qwen3.8"}
],
)
print(response.choices[0].message.content)The same call works with curl if you’d rather try it without installing anything:
curl -X POST https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions \
-H "Authorization: Bearer $DASHSCOPE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.8-max-preview",
"messages": [
{"role": "user", "content": "Explain what an open-weight model is"}
]
}'To confirm your account already has access to the preview, check the model field in the JSON response: it should return exactly qwen3.8-max-preview. If it comes back empty or with a different identifier, your account doesn’t have access enabled yet.
⚠️ Heads up: Qwen3.8-Max-Preview is, for now, a closed pay-per-token model. The open weights, according to the announcement, are coming “soon,” but Alibaba hasn’t given a confirmed date or license.
Impact and analysis
Alibaba’s pattern with Qwen3.8 repeats something already seen with DeepSeek, Kimi, and GLM: use a direct social media announcement, with immediate access to a preview version, to generate traction before publishing the technical paper or the weights. It’s an effective strategy for capturing curious developers, but it shifts the burden of verification onto the community, which has to wait for independent benchmarks to know whether the model delivers on its promise.
For developers in Latin America, the relevant point isn’t so much the parameter count but the compatibility with the OpenAI API: any project already using that format can point to Qwen3.8-Max-Preview by changing just the base_url and the key, without rewriting the integration code. That lowers the friction for comparing the model against what’s already used in production.
The most important limitation today is that there’s no way to independently verify the claim that Qwen3.8 is “second only to Fable 5.” There’s no results table on public benchmarks like Artificial Analysis or LMSYS Arena as of this article, and the announcement itself doesn’t include any comparison chart.
What’s next for Qwen3.8
There are three things to watch next: first, whether Alibaba publishes a model card with architecture, training data, and benchmark results; second, when the open weights appear on Hugging Face under the Qwen organization, where all previous versions were distributed; and third, whether independent labs incorporate Qwen3.8-Max-Preview into their public rankings to test Alibaba’s claim.
It’s also worth watching how the rest of the Chinese open model ecosystem reacts: if Qwen3.8 truly gets close to the frontier, DeepSeek, Moonshot AI, and Zhipu will likely respond with their own updates in the coming months, as has been happening throughout 2026.
📖 Summary on Telegram: See summary
Try it yourself: run the curl example above against Alibaba Cloud’s Token Plan and compare Qwen3.8-Max-Preview’s response with the model you already use in production.
Frequently Asked Questions
What is Qwen3.8?
It’s the latest language model in Alibaba’s Qwen family, announced on July 19, 2026 with 2.4 trillion parameters and plans to release it in open weights.
How many parameters does Qwen3.8 have?
According to Alibaba’s official announcement on X, 2.4 trillion. The figure is self-reported and doesn’t yet have a public model card detailing it.
Can you already use Qwen3.8?
Yes, through the preview version Qwen3.8-Max-Preview, available on Alibaba Cloud’s Token Plan, Qoder, and QoderWork.
When will the open weights for Qwen3.8 be released?
Alibaba said it would be “soon,” without an exact date or confirmed license as of this article.
Is it true that Qwen3.8 is the second best model in the world?
That’s a claim made by Alibaba in its own announcement. There are no independent benchmarks published yet to confirm or refute it.
Is Qwen3.8 compatible with the OpenAI API?
Yes, through DashScope’s compatible mode, the same mechanism used by other Alibaba Cloud models.
References
- Original post by @Alibaba_Qwen on X: the announcement of Qwen3.8 and Qwen3.8-Max-Preview.
- Qwen Cloud, international Token Plan: pricing for token-based access.
- Qianwen AI Platform, China Token Plan: pricing for access in the Chinese market.
- Qwen organization on Hugging Face: repository where the Qwen family’s open weights are published.
- QwenLM on GitHub: reference code and documentation for Qwen models.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion
Imagen destacada: Foto de Markus Spiske en Unsplash
0 Comments