Alibaba updates Qwen3.8-Max, leaves the -0902 checkpoint mapping unclear
Alibaba says the September 2nd update improves coding and Cowork performance; its Singapore service lists $2 per million input tokens and $6 per million output tokens.
By RuntimeWire Staff · Published
Primary source: Qwen / Alibaba
Why it matters
Alibaba documents a sparse model with lower per-token compute and a 2.4-trillion-parameter storage burden, while its Qwen3.8-Max service carries metered pricing. Until Alibaba maps the -0902 revision to downloadable weights and a managed endpoint, operators cannot safely transfer specifications or deployment assumptions to the update.
Alibaba's Qwen team announced an update to the Qwen3.8-Max model, calling the revision Qwen3.8-Max-0902 and saying further post-training targeted coding and Cowork tasks. The post also claims stronger performance across complex enterprise tasks and scientific research. It does not establish that a separate downloadable checkpoint was released that day.
RuntimeWire covered Alibaba's Commerce Agent Bench claims for Qwen3.8-Max on August 31st. The September 2nd post raises a separate deployment question: which open weights and managed endpoint, if any, correspond to the -0902 revision.
Alibaba's official post announces Qwen3.8-Max-0902 as a post-trained update, while the supplied documentation does not establish whether that name maps to the Qwen3.8-2.4T-A95B open-weight checkpoint.
That distinction matters because Alibaba uses three labels across its materials. The September post names Qwen3.8-Max-0902. The cloud documentation names Qwen3.8-2.4T-A95B. Qwen's official model card identifies Qwen3.8-2.4T-A95B as the open model underlying Qwen3.8-Max and lists 2.4 trillion total parameters, 95 billion activated parameters and a mixture-of-experts architecture. It does not identify that checkpoint as the -0902 revision or its managed equivalent.
Alibaba's public materials do not confirm whether corresponding -0902 weights became available or identify the endpoint serving that revision. Developers evaluating the update need to record the model identifier returned by the endpoint or attached to downloaded weights.
The identity gap changes the deployment math
For Qwen3.8-2.4T-A95B, Alibaba's model documentation specifies 2.4 trillion total parameters and about 95 billion activated parameters. Sparse routing selects part of the model for each step, reducing per-token computation relative to activating every parameter.
The complete checkpoint remains much larger than a 95-billion-parameter model. At 16-bit precision, 2.4 trillion weights alone require roughly 4.8 terabytes of memory, before key-value caches, routing, runtime memory and other overhead. Serving it also requires the weights to be distributed across accelerators, adding communication and latency costs.
Alibaba's documentation for Qwen3.8-2.4T-A95B lists 991,808 input tokens and up to 131,072 output tokens. The documented interface supports function calling, structured output, web search and context caching. Those specifications are useful for sizing agent workloads, though operators should not assign them automatically to Qwen3.8-Max-0902 while the identifier mapping remains unresolved.
The cloud prices belong to Qwen3.8-Max
Alibaba's model-pricing documentation lists Singapore pricing of $2 per million input tokens and $6 per million output tokens as of September 2nd. The documentation also lists lower prices for cache hits.
Those rates make the managed service a practical testing route for startups evaluating the model. Production agents can repeatedly incur input, output and cache charges as they inspect files, invoke tools, revise code and carry context across steps. Self-hosting requires enough accelerator memory and interconnect capacity for the full checkpoint even when sparse routing lowers the computation used for each token.
The pricing sharpens the model-identity issue. The model card connects Qwen3.8-2.4T-A95B to Qwen3.8-Max, but Alibaba's public materials do not say whether the priced service serves the -0902 revision. Operators therefore cannot assign the A95B model's documented limits and deployment behavior to the updated endpoint without checking the service configuration.
Alibaba's partnership roster identifies Jingren Zhou as an Alibaba partner and head of its Tongyi Foundation Model Division. Alibaba's leadership materials say he joined in 2016 after working as a Microsoft research-and-development partner and earned a computer science doctorate from Columbia University. Alibaba does not identify an individual founder for Qwen.
Alibaba co-founder and CEO Eddie Wu has placed Qwen inside a broader cloud push. RuntimeWire reported in August that QwenCloud combines models, agent tools and compatible APIs. Customers evaluating the update still need to identify which Qwen3.8 artifact an endpoint serves, measure its behavior on their workloads and apply the corresponding service prices.