क्वेन 3.8 2.4T

The open-source release of Qwen's latest flagship, and the first time a Qwen Max-class model has been published openly. Its sparse Mixture-of-Experts architecture holds 2.4T total parameters with roughly 95B activated per step, combining Gated DeltaNet and Gated Attention across 92 layers with a million-token context. It always reasons: every response begins with an internal reasoning pass before the final output.

This is the checkpoint the Qwen3.8-Max announcement promised. Alibaba used that post to say it would open-source the weights of a Max-class model for the first time, and the weights landed nine days later, on 12 August 2026, with no separate write-up. For a reader the practical meaning is straightforward: this is the same post-trained model behind the lab's hosted flagship, downloadable and runnable on hardware you choose rather than reachable only through one vendor's API. The licence is the caveat - it ships under a custom Qwen3.8-Max licence rather than the permissive terms Qwen's smaller open models carry, so commercial terms are worth reading before building on it.

The architecture is what makes a model this size servable at all. Alibaba describes 92 layers holding 2.4 trillion parameters with roughly 95 billion active per token, drawn from 512 experts of which ten routed plus one shared fire on any given token, and a hybrid backbone that runs Gated DeltaNet linear attention in three of every four layers with a full Gated Attention layer in the fourth. Multi-token prediction is trained in. Context is 262,144 tokens natively and extends past a million.

Alibaba is unusually explicit about what this checkpoint does not do, and the constraints are real rather than cosmetic. It is text only: multimodal input is not supported. Thinking cannot be disabled - every response begins with an internal reasoning pass before the final output, at any effort setting. The hosted sibling is the version that adds image input, a non-thinking mode and the million-token window by default, so the open weights are the reasoning core of the product rather than the whole product.

Because it is the same model, its reported numbers are the flagship's: Alibaba publishes 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro, 93.0 on PaperBench, 92.6 on GPQA Diamond, 43.6 on HLE, 82.8 on IFBench and 92.9 on the 256K eight-needle retrieval test. The same table shows it behind on DeepSWE v1.1 at 56.6 against 73.0 and 70.0, and on SWE-bench Pro against a leading 80.0.

Running it well takes some budgeting. The effort dial defaults to its highest setting, thinking from earlier turns is preserved by default - which Alibaba says helps agent consistency and cache reuse - and the lab's own recommendation inside a million-token context is to allow 262,144 tokens for reasoning and 131,072 for the final response. That split is the clearest statement of what this model is: one that spends a large share of its context thinking before it writes anything down.