Alibaba Qwen Team Open-Sources Qwen3.8-2.4T-A95B Weights, First Qwen-Max-Level Model Released

Key Highlights

Alibaba's Qwen team has truly pulled out all the stops this time. In August, the team officially released the weights of Qwen3.8-2.4T-A95B, marking the first time that a flagship model at the Qwen-Max level has been made fully public through open source, and it can be used for free. For the open-source community, this is equivalent to moving the top-tier capability that was once only accessible through paid APIs directly into a repository that anyone can download. The move also resets expectations for what open weights can deliver, because a model of this caliber was previously assumed to live behind a paywall indefinitely, available only to those willing to pay for hosted access. The release quickly drew attention from developers who had assumed that only closed models could reach this level, and early discussion focused on what self-hosting such a model would actually cost in practice rather than on whether it was merely possible.

Specific Capabilities and What Happened

According to a report from ITHome, the newly released Qwen3.8-2.4T-A95B has a total of 2.4 trillion parameters, but only 95 billion parameters are activated per token, which is a typical MoE, or mixture-of-experts, architecture. More importantly, it natively supports a context length of 262,144 tokens and can be further extended to 1,010,000 tokens, which means a super-long context of over one million tokens. This means it can read an entire book, a whole codebase, or even a large technical document in a single pass, without having to chop the input into pieces. The practical implication is significant: tasks that once required careful retrieval and chunking can now be handled in one continuous context window, which simplifies pipelines that were previously brittle, expensive, and hard to maintain across long documents. The naming itself is informative, because the A95B suffix signals the activated expert count, a convention that helps operators estimate the real compute needed before they download anything.

Technical Details

Looking at the parameter design, the combination of 2.4 trillion total parameters and 95 billion activated parameters reflects the idea of large model capacity plus low inference cost. The MoE architecture lets the model maintain a huge knowledge capacity while keeping the computing cost of each inference step within an acceptable range. The ability to extend to a million-level context relies on techniques such as chunked caching and sparse attention, making the training and inference of extremely long inputs feasible in both engineering and cost terms. For researchers, open weights also mean the model can be fine-tuned, distilled, and quantized, truly turning it into something they own. These techniques matter because they turn a theoretically interesting context length into something that can actually run without exhausting memory or blowing up latency on real hardware, which is where many long-context claims from the past ultimately fell short. For teams that cannot afford the full run, the open weights also unlock smaller derived models through distillation, so the 2.4T original can seed a family of cheaper siblings tailored to specific tasks.

Comparison with Competitors

Simply put, an open-source flagship has, for the first time, touched the threshold of closed-source top configurations. Previously, capabilities like Qwen-Max usually existed only in Alibaba Cloud's commercial APIs. Now that the weights are public, it directly benchmarks against the largest models in the open-source camp such as Llama and Mistral. Compared with dense models that require massive clusters to run, the MoE design of Qwen3.8 saves more video memory at equal effect, lowering the deployment barrier. Of course, the total scale of 2.4 trillion still requires considerable hardware, and ordinary developers would find it hard to run it completely locally. The honest caveat is that openness here means the weights are available, not that the model is light; running the full thing still demands serious infrastructure that only well-funded teams can provide, so the playing field is more level than flat. It is also a strategic signal: by open-sourcing the Max tier, Alibaba blurs the line that competitors use to justify premium pricing, pushing the whole market toward capability parity at the top end.

Industry Impact and Applicable Scenarios

For enterprises and developers, the greatest value of this release is control. Data-sensitive industries such as finance, healthcare, and government can deploy the model privately in their own server rooms, gaining flagship capability without sending their data beyond the boundary. For academic research, open weights mean reproducibility and auditability. Simply put, Qwen has pushed the open-sourcing of top-tier capability one step further, and has also made the boundary between open source and closed source increasingly blurred. The longer-term effect may be a slowdown in the willingness of enterprises to pay premium prices for closed models when an open equivalent is within reach, which pressures the entire commercial landscape to compete on service, latency, and support rather than on raw capability alone. In the near term, expect more enterprises to pilot private deployments and to compare open Qwen against closed APIs on their own workloads rather than on published leaderboards.