Alibaba’s Qwen team has released the weights for Qwen3.8-2.4T-A95B, the model it only gestured at when AI Insiders covered the company’s announcement post on 21 July. The card is live on Hugging Face now. It replaces the earlier promise with a full specification: parameter counts, architecture, benchmark scores, and three supported inference frameworks. A two-trillion-parameter target has become a model with a name that states its exact size.
The name itself carries the detail that matters most for anyone planning to serve it. Qwen3.8-2.4T-A95B holds 2.4 trillion parameters in total. Only 95 billion of them activate on any given token.
That split comes from a mixture-of-experts design built around 512 experts, of which 10 routed experts plus one shared expert fire per forward pass, according to the model card. The distinction determines cost. Storing 2.4 trillion parameters requires enough accelerator memory to hold the entire model, typically spread across many GPUs. The compute spent per token tracks the 95 billion active figure instead, closer to what a mid-sized dense model costs to run. Qwen is offering frontier-scale capability while keeping inference cost closer to what a mid-sized model would need.
The architecture builds on Qwen3.5’s foundation, per the card. It arranges 92 layers in a repeating pattern that mixes Gated DeltaNet blocks with Gated Attention blocks, each paired with a mixture-of-experts layer. Context length runs to 262,144 tokens natively. It extends to just over one million tokens, well beyond what earlier Qwen releases shipped with.
Qwen3.8-2.4T-A95B requires thinking mode for every response and cannot run without it. That constraint sits on the raw weights, not on the hosted product. Reasoning depth is adjustable through a reasoning_effort parameter set to xhigh, medium, or low, and a preserve_thinking option keeps prior reasoning context across turns. Qwen3.8-Max, the managed version sold through Qwen Cloud, adds vision input, an option to turn thinking off, and a one-million-token context window by default. Those features are absent from the open weights.
On the benchmark table Qwen published alongside the release, Qwen3.8-Max scores 86.6 on Terminal Bench 2.1 and 93.0 on PaperBench, both ahead of Qwen3.7-Max’s 74.5 and 64.8. The picture against outside models is mixed.
GPT 5.6 Sol scores higher on Terminal Bench 2.1, at 88.8. Qwen3.8-Max leads on PaperBench and on IFBench instead. Every one of those numbers comes from Qwen’s own evaluation harness, run primarily through Claude Code. The card includes no independent leaderboard result to check them against.
The card lists three serving engines by name: SGLang, vLLM, and TokenSpeed, each with a published integration recipe. That is the detail July’s announcement post could not offer. There were no weights yet to load into any of them.
Teams evaluating this class of open model can now size hardware against the 95 billion active parameters, not the full 2.4 trillion count. Three serving paths are ready to benchmark before any infrastructure commitment lands on a 2026 budget.
Alibaba’s Qwen team published the Qwen3.8-2.4T-A95B model card on Hugging Face on 13 August 2026.
