The Qwen3.8-Max was announced by Alibaba on August 3, 2026 as the most capable model in the Qwen family to date. It already appears in the QwenCloud catalog and can be called via API. The part that would allow you to download the parameters and try to run it outside the company's infrastructure, however, was not yet available in the announcement.
The Qwen team itself informed that the weights would be published “next week”, on Hugging Face and ModelScope, but did not indicate an exact date nor did it disclose, on the verified pages, the license that will accompany these files. Therefore, it is not correct to describe the Qwen3.8-Max as an open weight model already available on August 3. What exists now is an API-accessible service and a promise of future openness.
Qwen3.8-Max: available now versus promised
| Item checked | Available August 3, 2026 | Promised or pending | Editorial verdict |
|---|---|---|---|
| Model access | API on QwenCloud, with the identifier qwen3.8-max | — | Available as a hosted service |
| Model weights | Not published at time of verification | Publishing on Hugging Face and ModelScope “next week” | It is not yet possible to download and host the official weights |
| Weight license | Not reported on official verified pages | Must be checked when the repository is published | Rights of use, modification and redistribution cannot yet be assessed |
| Own execution | Not available through official public weights | May be technically possible after publication, depending on files and license | API does not equate to self-hosting |
| Context | Total announced window of 1 million tokens; catalog reports up to 991 thousand entry tokens | Limits may change depending on endpoint and configuration | Available, but subject to service rules |
| Modalities | Text, image and video as input; text as output | Other features will depend on the documentation and integrated tools | Multimodality confirmed in the catalogue |
| API Price | $2 per million input tokens and $6 per million output tokens at the consulted list price | Values, discounts and plans may change | Price published but volatile |
| Performance | There are results released by the manufacturer and starting positions in Arena | Broader, more stable independent testing | Promising but not conclusive |
Available API does not mean open model
An API allows you to send data to a model running on the vendor's servers. The user receives the answers, but not necessarily the internal files, training code, parameters or the right to modify and redistribute the system.
Open weights represent another level of access: parameters learned during training are made available for download under certain conditions. Even so, open weights are not automatically synonymous with fully open source artificial intelligence.
According to the criteria of the Open Source Initiative, the evaluation of an AI system involves not only the parameters, but also terms that ensure freedom of use, study, modification and sharing, in addition to the information and code necessary to understand how the system was produced. This is a specific and debated definition in the industry, but it is useful to avoid “open” being used as a generic expression.
In the case of Qwen3.8-Max, there are three stages that should not be confused:
- Accessible API: confirmed on August 3rd.
- Weights available for download: promised, but still pending.
- License and actual degree of openness: still need to be examined.
Even if the weights are published on schedule, you will need to read the license file and check for any restrictions on commercial use, redistribution, creation of derivative models, user limits, or attribution obligations.
What the technical sheet confirms
2.4 trillion parameters, with 95 billion active
Alibaba states that Qwen3.8-Max has 2.4 trillion total parameters and uses 95 billion per request. The model adopts a sparse expert mixing architecture, known as Mixture of Experts or MoE. Instead of activating the entire system for each token, the architecture selects specialized parts of the model during inference.
This design can reduce the computation required per response when compared to a dense model of the same overall size. It does not mean, however, that Qwen3.8-Max has the same storage or deployment cost as a model with only 95 billion parameters.
As a rough estimate, 2.4 trillion parameters would take up about 4.8 terabytes in BF16, 2.4 terabytes in 8-bit, or 1.2 terabyte in 4-bit, before metadata, cache, execution structures, and other overhead. The final files, formats and quantizations still need to be known. Therefore, “available weights” should not mean “model that is easy to run on a regular computer.”
Reuters also noted that the total number puts the model close to Moonshot AI's Kimi K3, cited with 2.8 trillion parameters. The count alone does not prove superior quality: architecture, data, post-training, inference method and evaluation conditions also influence the result.
Context, modalities and limits
The official catalog features a total context window of 1 million tokens, with up to 991 thousand input tokens and approximately 131 thousand output tokens. For requests with reasoning, the page reports up to 983 thousand input tokens and a maximum reasoning limit of 262 thousand tokens.
The model accepts text, images, and videos as input and produces text as output. The ad also highlights extensive document analysis, visual content and long videos, but examples provided by the manufacturer itself should be treated as promotional demonstrations until they are reproduced under independent conditions.
QwenCloud reports compatibility with interfaces inspired by OpenAI and Anthropic API specifications. This can make it easier to adapt existing applications and programming wizards, but protocol compatibility does not guarantee identical behavior between vendors.
How much does it cost to use Qwen3.8-Max
In the consultation carried out on August 3, the specific model page showed the following list price:
- Input: $2 per million tokens;
- Output: $6 per million tokens;
- Implicit cache input: $0.25 per million;
- Explicit cache read: $0.17 per million;
- Explicit cache creation: $2.50 per million.
These values refer to the QwenCloud service and do not anticipate the cost of running future weights on our own infrastructure. The actual price of an application will also depend on the input to output ratio, reasoning usage, context size, tool calls, cache, and number of attempts required to complete a task.
As prices and plans are changeable information, the official page must be checked again immediately before publication or any commercial comparison.
What performance claims really show
Alibaba reported that Qwen3.8-Max ranked fifth in Text Arena, second in Vision Arena and fourth in a ranking focused on interface development. Reuters also recorded the model's prominent position among Chinese offerings.
The Arena changelog states that Qwen3.8-Max was added to the text, vision, and code leaderboards on August 2, 2026, the day before the official announcement. Newly added models may have fewer comparisons and larger uncertainty intervals, while positions change as new votes are collected.
This makes the result a relevant signal, but not a definitive conclusion. Arena measures preferences in comparisons between answers and can capture aspects of usefulness and style that static benchmarks do not measure. At the same time, the position is not a substitute for specific assessments of accuracy, security, cost, latency, languages, tool usage, or reliability in extended tasks.
The tests described in the announcement — including multi-day programming projects, professional analysis, and agent operations — were organized or presented by the Qwen team itself. They can guide evaluation hypotheses, but should not be reproduced as independent proof of superiority.
What will need to be checked when the weights are published
Updating the article should not be limited to changing “promised” for “available”. The publication needs to be audited in at least six points:
- Repository Source: confirm that the page belongs to the official Qwen organization on Hugging Face or ModelScope.
- License: identify permissions and restrictions for use, modification, redistribution, and commercial applications.
- Published files: check if they are full weights, quantizations, checkpoints or just auxiliary configurations.
- Technical documentation: browse architecture, tokenizer, software requirements, formats, and inference statements.
- Infrastructure requirements: check memory, number of accelerators, parallelism and support in tools like vLLM or SGLang.
- Reproduction of results: separate third-party executable benchmarks from manufacturer-controlled demonstrations.
You will also need to check whether the published version exactly matches the qwen3.8-max endpoint or whether there are differences between the hosted edition, a preview version, and the distributed weights.
So Qwen3.8-Max is open source?
As of August 3, 2026, not yet. Qwen3.8-Max was available as a service via API, with pricing, terms and limits published on QwenCloud. The official weights, however, remained scheduled for the following week, and the license could not yet be evaluated.
The most accurate formulation at this point is: the Qwen3.8-Max is a hosted model whose manufacturer has promised to publish the weights. After publication, it can be described as an open weights model only if the files are effectively accessible under terms that permit this use.
For technical teams, the API already allows prototyping, measuring quality and estimating costs. For those who intend to host, adapt or audit the model outside of the Alibaba service, the decision still depends on the weights, license and documentation.
Save this comparison and check out the update when the official repositories are published. The decisive change will not be a new announcement, but the verifiable availability of the files and their terms of use.
