Qwen3.8-Max Open Weights Make a 2.4T Model Downloadable

The First Max-Class Qwen You Can Download

On August 12, 2026, the Qwen team published the weights for Qwen3.8-2.4T-A95B on Hugging Face, making it the first model in Alibaba's Max tier to be released as an open-weight download rather than an API-only endpoint. When we covered the Qwen3.8-Max API launch on August 4, the weights were promised but not yet available. They are now live, alongside an FP8 variant published the same day.

• 2.4 trillion total parameters with 95 billion activated per token, published under the qwen3.8-max license
• 262,144-token native context, extensible to roughly 1,010,000 tokens
• 512 experts across 92 layers, with 10 routed plus 1 shared expert active per token
• Text-only, with an official FP8 checkpoint shipping alongside the BF16 weights

The distinction between a published benchmark and a downloadable checkpoint is not a small one. An API endpoint can be measured; a weight file can be inspected, fine-tuned, distilled, quantized, and audited. Moving a Max-class model across that line changes what independent researchers are able to do with it.

What Makes the Open Weights Different From the API Version?

The downloadable checkpoint and the hosted Qwen3.8-Max service are not identical, and the model card is explicit about it: Qwen3.8-2.4T-A95B is a text-only model, while the API tier also handles vision. The open release also requires thinking mode for all interactions rather than exposing the full low/medium/high/xhigh spread the hosted product offers.

One widely repeated claim is worth correcting. Several secondary write-ups have described the open weights as having lost the long-context capability. The official model card does not support that reading — it lists 262,144 tokens natively with extension to approximately 1,010,000. Where a vendor's own model card and downstream commentary disagree on a specification, the model card is the document to trust.

The companion Qwen3.8-27B, announced at the same time as the Max tier, has not yet appeared on the official Qwen organization. The Qwen3.8-27B repositories currently visible on Hugging Face are community uploads rather than first-party releases, so anyone waiting on the single-GPU member of this generation is still waiting.

Inside the Hybrid Attention and MoE Architecture

The architecture is the more interesting story for anyone tracking how frontier models are being built in 2026. Qwen3.8-2.4T-A95B alternates between two kinds of layers in a repeating block: three Gated DeltaNet stages feeding mixture-of-experts blocks, followed by one gated full-attention stage feeding another MoE block, repeated 23 times across a hidden dimension of 8192.

That alternation is a deliberate trade. Full-attention layers preserve exact token-to-token relationships but scale quadratically with sequence length. Linear-attention layers such as Gated DeltaNet compress history into a bounded recurrent state, which keeps memory flat as context grows but discards some precision. Interleaving them at a 3:1 ratio keeps most of the long-context cost linear while retaining periodic full-attention passes. If the sparse-activation side of this design is unfamiliar, our explainer on sparse mixture of experts covers why activating 95B of 2.4T parameters is the economics that makes a model this size servable at all.

On the published evaluations in the model card, the model reports 92.6 on GPQA Diamond, 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro, and 56.6 on DeepSWE 1.1. As always, vendor-reported numbers are a starting point for your own evaluation rather than a substitute for it.

What Hardware Does a 2.4T Model Actually Need?

This is not a laptop model, and it is worth being straightforward about that. NVIDIA published a same-day deployment guide describing the model running on a GB300 NVL72 — 72 Blackwell Ultra GPUs linked by 130 TB/s of NVLink bandwidth — where it sustains over 4,000 tokens per second per GPU and over 350 tokens per second per user at FP8 precision. NVFP4 support is described as a further optimization still to come.

For deployment, SGLang, vLLM, and NVIDIA Dynamo are the recommended open-source paths, with NIM available as a container option. Community quantizations have already begun appearing, which is precisely the pattern open weights enable and closed endpoints do not.

Why Open Max-Class Weights Matter

The practical significance is less about who can run 2.4 trillion parameters at home and more about what becomes possible downstream. Open weights at this scale give organizations with on-premise or sovereignty requirements a frontier-class option that never leaves their infrastructure. They also give the research community a teacher model for distillation — the mechanism behind most of the capable small models that do run on consumer hardware.

That lineage is visible across the past month. Kimi K3's 2.8T open-weight release established that models at this scale would be published openly at all; Qwen3.8-Max confirms it was not a one-off. More coverage of open-weight releases lives in our artificial intelligence section.

Sources: Qwen3.8-2.4T-A95B model card, Hugging Face — August 12, 2026; NVIDIA Technical Blog — August 12, 2026.