Qwen3.8 Max Open Weights: 2.4T Parameters and Context

What went up on Hugging Face

Alibaba's Qwen team uploaded Qwen3.8-2.4T-A95B to Hugging Face on August 12, 2026: 2.4 trillion parameters in total with 95 billion active per token, arranged as a mixture-of-experts model with 92 layers, an 8,192 hidden dimension, and 512 total experts of which 10 routed plus 1 shared fire on any given token. It combines Gated DeltaNet and Gated Attention, natively handles 262,144 tokens of context extensible to 1,010,000, and runs text-only with thinking mode always on, prefixing every response with reasoning wrapped in think tags whose depth is set by a reasoning_effort parameter. [1]

The companion GitHub repository, updated the same day, describes the release as bringing a Qwen-Max-class model to open release for the first time, with weights and code under Apache 2.0 and a separate model license posted alongside the files on Hugging Face and ModelScope. A smaller companion model, Qwen3.8-27B, went open-weight around the same time. [2][3]

The API came first, and it was multimodal

Qwen3.8-Max reached general availability nine days earlier, on August 3, through Alibaba's own hosted Model Studio API, OpenAI-compatible and DashScope-compatible, with the open weights announced as coming the following week. The hosted flagship accepts text, image and video as input and returns text, with a model page listing a 1 million token context window and a 991,000 token maximum input. [3]

Atlas interpretation: The checkpoint that landed on Hugging Face on August 12 is not that same multimodal service. It is text-only, with vision and video handled by the hosted product rather than the self-hostable weights. Read against the hosted launch, the open-weights release looks like the core reasoning model Alibaba is willing to hand over, stripped of the perception stack it sells access to through the API. [1][3]

Sixteen days behind Kimi K3, and priced like it

Moonshot AI had put a larger set of open weights on Hugging Face sixteen days earlier: Kimi K3, at 2.8 trillion parameters, released July 27, 2026, with native visual understanding and its own million-token context window. Qwen3.8's 2.4 trillion parameters made it the second largest open-weight release in the world rather than the largest, a distinction that lasted from July 27 until August 12 and then passed to Alibaba only in a qualified sense: bigger than everything except the model that had just displaced it. [4]

Days before the open weights shipped, TheNextWeb reported that Alibaba intended to charge its largest commercial users of the model a share of revenue once they crossed a threshold, while leaving the weights themselves free to download and run. The article described this as Alibaba mirroring Moonshot, whose Kimi K3 license asks partners for revenue sharing, up to 30 percent, once a partner's annual sales pass $20 million; Alibaba's own percentage was not yet fixed at the time of that report. [5]

Atlas interpretation: That the two labs converged on the same mechanism within weeks of each other is the more durable part of the story. "Open weights" for a frontier-scale Chinese model in mid-2026 had stopped meaning unconditional; it meant free to a point, with a commercial agreement waiting on the other side of that point for the users large enough to matter. Qwen3.8's own terms, tied to Alibaba's separate model license rather than Apache 2.0, sit inside that same pattern rather than outside it. [5][2]