Qwen3.8-Max has crossed a more concrete threshold. On August 12, the ModelScope repository for Qwen/Qwen3.8-2.4T-A95B showed a published model card with post-trained weights and configuration files in Hugging Face Transformers format.
The repository names vLLM, SGLang and TokenSpeed as compatible serving paths, and ModelScope metadata shows a storage size of about 4.89TB. Alibaba had already previewed Qwen3.8-Max through Qoder and Token Plan; the new event is the availability of a Max-class open-weight artifact.
The weight is the new checkpoint
“This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.”
The model card lists 2.4T total parameters, 95B activated parameters, 262,144 native context tokens and an extension path to 1,010,000 tokens. The architecture uses 92 layers, 512 experts, 10 routed experts plus one shared expert, and a 248,320-token padded vocabulary.
It also draws a boundary with the managed Qwen3.8-Max service. The open-weight model is text-only and requires thinking mode. Qwen Cloud’s official Qwen3.8-Max service adds features such as vision input, non-thinking mode, default 1M context and built-in tools.
Open weights change who can audit it
For developers, the difference between an API discount and open weights is control. Discounts can expire. A downloadable model lets teams test serving engines, quantization, private codebases, long documents and security boundaries without routing every workload through a public API.
The 4.89TB size also keeps the claim grounded. This is not a casual laptop model. Most users will encounter it through hosted inference, quantized community builds or enterprise clusters. The open release mostly expands the number of people who can inspect the model, reproduce scores and challenge Alibaba’s claims.
The benchmark table needs footnotes
Official scores put Qwen3.8-Max at 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro, 93.0 on PaperBench and 86.1 on OSWorld-Verified. These figures place it in serious coding-agent and cowork territory, but the footnotes matter: harnesses, timeouts, context windows and baseline sources differ across tests.
The cautious reading is that Qwen3.8 is now worth independent replication, not that one table proves blanket superiority. The next checks are license clarity, stable downloads, vLLM and SGLang recipes, third-party benchmark runs and how much capability quantized releases keep.
Sources: ModelScope Qwen3.8-2.4T-A95B model card, Qwen3.8-Max official blog, Qwen Cloud model documentation, DataCamp, CocoLoop; verified open-weight status, repository size, total and active parameter counts, context length, framework compatibility, thinking-mode limits, benchmark scores and managed-API differences.