Qwen 3.8 Open Weights: What You Actually Get

The Qwen 3.8 open weights are the files for Qwen3.8-2.4T-A95B, a text-only model that always reasons before it answers and natively handles 262,144 tokens. Qwen, the AI team at Alibaba, publishes them on Hugging Face.

The hosted Qwen3.8-Max is built on the same base model and adds image input, a non-thinking mode and a 1M default context. The download also comes in two sizes and carries a custom license instead of Apache 2.0.

What the Qwen 3.8 open weights contain

Alibaba's Qwen team calls Qwen3.8-2.4T-A95B its first Qwen-Max-class model released as downloadable files, and only 95B of its 2.4T parameters work on any one token. The "2.4T" counts every parameter stored in the checkpoint, while "95B" counts the parameters a single token passes through, about 4% of the total. Any account can download the repository, because it is public and not gated, and the weights come in the Transformers safetensors format.

The saving comes from a mixture-of-experts design with 92 layers and 512 experts, where each token is routed to 10 of them plus 1 shared expert. The other experts stay idle for that token, which keeps the work per token far below what a dense 2.4T model would need. Storage is a different matter, since every expert still has to be loaded.

Qwen 3.8 open weights vs hosted Max: what is missing

Only text goes in and the reasoning step always runs, while the official version adds image input, a non-thinking mode and a 1M default context. The model card names that official version Qwen3.8-Max, built on Qwen3.8-2.4T-A95B with more features. It also lists official built-in tools among the additions, so the downloaded checkpoint is a subset of what the hosted product offers.

The card states that the open model requires thinking mode for all interactions and that thinking cannot be disabled. The reasoning_effort setting changes how deep the reasoning goes, with xhigh as the default and medium and low as the other values, but none of the values turns it off. A task that needs plain, immediate replies has to use the hosted model.

Context length differs the same way. The checkpoint handles 262,144 tokens natively, and the card lists about 1,010,000 tokens as the extended ceiling. A project that sizes memory on the 1M figure is sizing it on the hosted product, so 262,144 is the safe planning number for the download.

Which Qwen 3.8 download fits: BF16 or official FP8

Both repositories are far beyond a single machine: the full-precision files take 4.45 TiB and Qwen's own 8-bit build takes 2.27 TiB, about half. The full BF16 repository is split into 213 safetensors files. The FP8 repository, named Qwen3.8-2.4T-A95B-FP8, is published under Qwen's own account and carries the same "other" license tag, so quantizing to FP8 leaves the terms unchanged.

Serving either build takes several machines working together. The model card lists vLLM, SGLang and TokenSpeed as compatible engines and recommends dedicated serving engines for production or high-throughput work. For a single workstation, the smaller sibling is the better route, and Qwen3.8-27B VRAM: 7GB to 56GB by Quant Level covers the 27B model for local hardware.

Is the Qwen 3.8 open weights license Apache 2.0?

No: Hugging Face tags the repository "other", and two scale thresholds add duties that the standard Apache terms never ask for. The license file is titled Qwen3.8-Max License. Its grant is broad, covering use, copying, modification, merging, redistribution, sublicensing, sale, deployment, hosting, fine-tuning and derivative works, all free of charge.

The first threshold applies to a commercial product or service with more than 100,000,000 monthly active users or US$20,000,000 in monthly revenue. That product must show the model name prominently in its user interface.

The second applies to a business that offers Model as a Service or an AI Work Assistant. If its combined revenue passes US$50,000,000 in any 12 consecutive months, it needs a separate license from Qwen before any commercial use.

Internal use is exempt from the second condition, as long as the software, its outputs and its model capabilities are not made available to any third party. The same terms apply to the official FP8 repository.

Why the Qwen 3.8 benchmark table is not a ranking

Qwen's footnotes state that every score comes from its own testing and that the harness changes from model to model, so the rows measure different setups. The model card puts Qwen3.8-Max next to other models in one comparison table. The footnotes then list how each row was run, and the setups differ in agent, timeout and scoring rule.

SkillsBench uses the public v1.1 set of 87 tasks and averages three runs per task. Opus 4.8 and Fable 5 run on Claude Code, GPT-5.6 Sol runs on Codex, and the Qwen models run on OpenCode.

Terminal Bench 2.1 is run by Qwen as avg@10 with a 5-hour timeout and max_tokens 131,072, while for every other model Qwen reports the best score published under any setup. The card adds that Fable5 results may involve fallbacks.

The card's footnotes are the only place that spells these differences out. They sit on the Hugging Face model card page, which is long and full of tables. To read that page as Markdown text, or to paste it into an AI assistant, convert the web page to Markdown first.

Summary: two products share the Qwen 3.8 name

The files anyone can download and the API-served model start from the same base weights, yet they differ in input types, reasoning control, context length and terms. The download is Qwen3.8-2.4T-A95B: text-only, always reasoning, and natively limited to 262,144 tokens. Image input, a non-thinking mode and the 1M default context belong to Qwen3.8-Max.

The files weigh 4.45 TiB in BF16 or 2.27 TiB in Qwen's FP8 build, and both need multi-machine serving. The license is not Apache 2.0, and its duties start at 100,000,000 monthly active users, US$20,000,000 in monthly revenue, or US$50,000,000 over 12 months for Model as a Service and AI Work Assistant businesses.

A team that needs image input or plain replies should use the hosted model, and a team with multi-machine serving and a text-only workload can download the files. A team with one workstation should look at the 27B model instead. Every benchmark score is Qwen's own result, so each claim about "Qwen 3.8" should be matched to the product it describes.