The open-weight release of Alibaba's Qwen3.8-Max flagship — a 2.4-trillion-parameter sparse mixture-of-experts model with roughly 95 billion parameters active per token, downloadable from Hugging Face and ModelScope in August 2026. The released checkpoint is text-only and runs in thinking mode.
Last verified · 2026-08-12 · by Moe Ameen
Qwen3.8-2.4T-A95B is the downloadable, open-weight release of Qwen3.8-Max, the flagship model from Alibaba's Qwen team. It arrived on Hugging Face and ModelScope in August 2026, some days after the model was made accessible through Qwen Chat and the Qwen API — this is the version whose actual weights you can pull and run yourself. It is a sparse mixture-of-experts (MoE) model with roughly 2.4 trillion total parameters and about 95 billion active per token, built on the architectural foundation of Qwen3.5. Per the official model card, the released checkpoint is a text-only model and requires thinking mode for all interactions — worth flagging, because earlier announcement coverage described the broader Qwen3.8 line as multimodal; the open-weight artifact on Hugging Face is text in, text out.
The architecture is what keeps a model this large economical to serve. Only a fraction of the 2.4T parameters activate on any token: a learned router picks a handful of experts (the design routes across a large expert pool with a shared expert plus a small set of routed experts per token), so serving cost tracks the ~95B active count rather than the full 2.4T. It interleaves linear-attention "Gated DeltaNet" layers with standard gated-attention layers — the same hybrid-attention lineage as Qwen3.5 — which helps it handle a very long context: 262,144 tokens natively, extensible toward roughly one million, with output length up to about 128K. It also exposes configurable reasoning depth (reported as low/high/xhigh tiers), so you can trade compute for answer quality per request. On Alibaba's own benchmark reporting it scores strongly on research and reasoning suites — figures in the low-90s on PaperBench and GPQA Diamond among them — and is positioned as competitive with leading frontier models; treat those as vendor numbers until independent evaluations land.
The license is the twist. The weights ship under a custom "qwen3.8-max" license rather than the permissive Apache 2.0 terms used on smaller Qwen releases: you can download and run the model, but Alibaba has signaled revenue-sharing terms for large commercial users and cloud providers that offer it as a service at scale — the specific thresholds and percentages were still being finalized at the time of writing. And a 2.4T-parameter model is not casual hardware: practical serving is a data-center job (reference deployments use rack-scale Blackwell-class GPU systems), though FP8, GGUF, and NVFP4 quantized builds and serving paths like SGLang, vLLM, and NVIDIA NIM lower the barrier. If a spec, the license terms, or the exact release date matters to your build, confirm it on the official Qwen model card before you commit.
The news in this release is control: the weights are downloadable, so a team can run Qwen3.8-Max on its own hardware and keep every prompt in-house. That is a real reason to choose it — but it changes nothing about what the model produces, which is text and only text. The open-weight checkpoint renders no video frame, designs no image, sizes nothing for a feed, holds no brand voice across a week of posts, and publishes to no platform. Self-hosting a 2.4T-parameter brain gives you private, frontier-grade drafting; it does not give you a content operation. Kompozy is the layer that turns that private draft into the finished, on-brand feed. Point your self-hosted Qwen endpoint at Kompozy through bring-your-own-key on the Founding tier, and the drafts and angles it reasons out become Photo Posts, brand-exact Carousels via HyperFrames, Quote Graphics, Persona and HeyGen avatar video, Clipped Shorts, plus native Text Posts, a Blog Article, and an Email Newsletter — none of which a raw text model can make.
The pairing is efficient precisely because the two halves don't overlap. Run Qwen3.8-2.4T-A95B where its scale earns its keep — long-context reasoning over messy source material, mined privately on your own boxes — and let its thinking mode surface the ten sharpest angles. Then hand each to Kompozy, where your Persona Brief and banned-word filters cast the text into your actual brand voice, reframe every output to 9:16, 1:1, and 16:9, burn in branded captions, and schedule and publish the whole set across nine destinations — the eight social platforms plus blog and email — with Autopilot and a per-post review pipeline. For a team that adopted the open weights for cost or data-governance reasons, this keeps the entire generation stack under your control right up to the render-and-ship boundary, which is exactly the boundary Kompozy owns. The model thinks privately; Kompozy makes it look like your brand and distributes it everywhere.
It is the open-weight release of Alibaba's Qwen3.8-Max flagship, published on Hugging Face and ModelScope in August 2026. It is a sparse mixture-of-experts model with roughly 2.4 trillion total parameters and about 95 billion active per token, built on the Qwen3.5 architecture. Per the official model card, the released checkpoint is text-only and runs in thinking mode.
The weights are downloadable, but not under a fully permissive license. They ship under a custom "qwen3.8-max" license rather than Apache 2.0, and Alibaba has signaled revenue-sharing terms for large commercial users and cloud providers offering it as a service at scale; the exact thresholds and rates were being finalized at the time of writing. Confirm the license on the official model card before commercial use.
The released open-weight checkpoint is text-only, per the official Hugging Face model card — text in, text out, in thinking mode. Some announcement coverage described the broader Qwen3.8 line as multimodal, but the downloadable artifact documented here does not accept images or video. Verify on the model card for your specific build.
A 2.4-trillion-parameter model is a data-center workload; reference deployments use rack-scale Blackwell-class GPU systems. Because only ~95B parameters are active per token, serving cost tracks the active count, not the full 2.4T, and FP8, GGUF, and NVFP4 quantized builds plus serving stacks like SGLang, vLLM, and NVIDIA NIM lower the barrier — but this is not a run-it-on-a-laptop model.
No. It generates and reasons over text but produces no video, images, or designs, enforces no brand voice, and publishes to no platform. To turn its drafts into finished, on-brand posts across platforms, pair it with a content engine like Kompozy — which can call your own Qwen key on the Founding tier — that renders the media and handles scheduling and publishing.