On August 3, 2026, Beijing time, Quantum Bit publicly released information about the new entry of Alibaba's Qwen3.8 through its information flow. The official product page of QwenCloud simultaneously displayed qwen3.8-max, model trials, pricing, and API call entries. The copyright holder for this page is Intelligent Cloud Computing (Singapore) Private Limited; however, the official page does not provide a complete availability range for all countries, regions, accounts, and packages. It also fails to list the partition launch times and account opening conditions, with only the public entry and listed items on the page currently confirmed.
Hybrid Expert Architecture and Long Context
QwenCloud categorizes Qwen3.8-Max as a hybrid expert model with a total parameter count of two point four trillion, featuring product tags for inference, visual understanding, and text generation. The page indicates that inputs can be images, text, and video, while outputs are listed as text. Two point four trillion is the total number of parameters and should not be interpreted as all parameters being activated per request; at the time, the active parameter count was not disclosed on the page. Hybrid expert here refers to the official classification of architecture, with no details provided on the public page regarding expert routing rules or the scale activated per request.
According to the specifications table on the official product page, Qwen3.8-Max has a total context length of one million tokens, with maximum input and output in normal mode being ninety-nine point one thousand and thirteen point one thousand tokens respectively. The maximum input for thinking mode is ninety-eight point three thousand tokens, with an additional listing of twenty-six point two thousand tokens as the maximum inference specification on the page. These values represent the hosted product limitations at the time of access and do not commit to specifications for local deployment or weights that have yet to be announced. Normal and thinking modes are listed separately in the table, so the upper limits cannot be added together to form a single request allowance.
API Pricing, Caching and Feature List
In terms of pricing, QwenCloud listed a rate of two dollars per million input tokens and six dollars per million output tokens at the time. The implicit cache input was priced at zero point two five dollars per million tokens. Explicit cache creation and read operations were priced at two point five dollars and zero point one seven dollars per million tokens respectively. The page also listed limits of two million tokens per minute and fifteen thousand requests per minute; these prices and limits are based on the values seen when accessing the page on August 3rd, with no guarantee of permanence. The three types of caching pricing correspond to different billing items, so this set of numbers cannot be combined into a fixed total cost per call.
The feature list includes prefix continuation, function calling, context caching, structured output, batch processing, web search, and fine-tuning. QwenCloud explains that context caching is used to save shared prefixes for long requests in order to reduce redundant calculations, improve latency, and lower costs. This is the official description of the product mechanism; however, the selected materials do not provide independent measurements of latency, cost, or hit rates under different request sizes. The feature names can only confirm what is listed on the page and cannot independently prove that they have been performance-verified across various loads and edge cases.
Built-in Tools and Official Capability Claims
The product page lists five built-in tools for the Responses API: code_interpreter, web_extractor, web_search, t2i_search, and i2i_search. The DashScope example on the page uses MultiModalConversation.call with the model identifier qwen3.8-max, where a single message simultaneously inputs an image URL and a text question. The example uses the international API address of DashScope, but it is not sufficient to confirm the actual availability for each region, account type, or plan. Tool names and code samples provide clues at the interface level; however, the page does not offer success rates or error rates for tool tasks in the same section.
QwenCloud positions its product around long-term planning, programming, professional office work, and native multimodal understanding, claiming that the model can handle autonomous programming projects spanning more than ten days, as well as tasks related to law, finance, and design. It also claims the ability to analyze extremely lengthy documents and videos. These capabilities are self-reported by the company; the selected materials do not provide reproducible project trajectories, original evaluation results, or third-party tests. Therefore, it cannot be inferred that the model has passed professional certifications in legal, financial, or design fields. The product positioning and reproducible results represent two different types of evidence, with the current public page only supporting the former.
Weight Preview and Unreleased Records
Quantum Bit provides a grade assessment of Qwen3.8's overall performance based on publicly available information, while the official page mainly offers features, specifications, and self-reported capabilities from the company. The currently selected materials do not include third-party benchmark methods, original results, or reproducibility records; therefore, this grade assessment is not considered an independently confirmed conclusion. Quantum Bit also announced that Max weights will be released next week along with Qwen3.8-27B; however, at the time of writing, the official product page did not provide specific dates, repository addresses, or licenses. Media announcements provided a subsequent timeline but do not guarantee that corresponding documents have been made public or that licensing terms are finalized.
The existing publicly available information does not disclose Qwen3.8-Max's training data sources, active parameter count, training compute volume, hardware configuration, energy consumption, security evaluation details, or third-party audits; the complete regional availability is also not listed. Subsequent verifiable records will include official weight files, licenses, technical reports, security evaluations, and independently reproducible tests. Until these records are made public, the release scope is limited to what was visible on the hosted product page and media announcements at that time, without considering the weight release plan as a completed release. If the company subsequently changes pricing, limits, or product status, updates should be based on new publicly available records rather than continuing with values from August 3rd.