AlibabaのQwen3.8-Maxは2.4兆パラメータのオープンウェイトMoEモデル——その実態と活用法Qwen3.8-Max Is Alibaba's Biggest Open-Weight Bet Yet. Here's What You

Alibaba has announced Qwen3.8-Max, a 2.4-trillion-parameter Mixture-of-Experts (MoE) model that the company says will be released under an open-weight license next week. According to the source, the announcement was made on August 3, 2026, and it matters because a model at this scale becoming freely downloadable would place one of the largest publicly available language models in the hands of developers, researchers, and self-hosters who have so far leaned on proprietary APIs for frontier-level capability. The accompanying blog post breaks down the official benchmark table, clarifies what "open-weight" actually means in this case, and walks through how to run the model through Claude Code.

The headline figure is the parameter count. At 2.4 trillion total parameters, Qwen3.8-Max sits well above most previously released open models. It is worth noting that the number reflects the MoE architecture rather than a dense network. In a Mixture-of-Experts design, the model is divided into many specialized subnetworks, or "experts," and a routing mechanism activates only a subset of them for any given token. This means the count of parameters that are actually computed per forward pass is typically far smaller than the total, which is how MoE models aim to deliver strong quality while keeping inference costs more manageable than a dense model of comparable size would require. Anyone planning to self-host should keep this distinction in mind, because the storage and memory footprint still scales with the full parameter set even when compute per token does not.

The term "open-weight" is another point the article addresses directly, and the distinction is important. Open-weight generally means the trained model weights are made available for download and local use, but it does not automatically imply that the training data, training code, or a fully permissive open-source license are included. Readers evaluating the release should therefore check the specific license terms, any usage restrictions, and whether commercial deployment is permitted, since these details vary considerably between labs and even between successive model releases. The blog's focus on defining the phrase suggests the release is likely to sit somewhere on the spectrum between fully open source and a restricted community license, as has been common across the industry.

For practical use, the article describes running Qwen3.8-Max through Claude Code, Anthropic's command-line coding tool. Claude Code is normally associated with Anthropic's own Claude models, but tooling of this kind can often be pointed at alternative model endpoints, which appears to be the workflow the post demonstrates. Using an open-weight model inside an established coding agent is an increasingly common pattern, as it lets developers keep a familiar interface while swapping the underlying model for one they can host themselves or access at lower cost. Whether the integration relies on an API-compatible endpoint or a local server is the kind of detail worth confirming against the original instructions before attempting to reproduce it.

The wider context helps explain why a release like this draws attention. The Qwen family is Alibaba's line of large language models, and earlier Qwen releases have been widely adopted precisely because many of them were distributed with downloadable weights. A steady stream of Chinese research labs, including the teams behind Qwen and DeepSeek, has pushed capable open-weight models into public circulation, applying competitive pressure on both closed frontier providers and other open-model efforts. Each new open release also feeds an ecosystem of adjacent tooling, from quantization projects that shrink models for consumer hardware to inference servers and agent frameworks that make deployment more approachable.

As with any pre-release announcement, some caution is warranted until the weights and full documentation are actually published. Benchmark tables provided by a model's creator reflect the vendor's own evaluation setup, and independent testing typically follows once a model is broadly available. The scale, the open-weight framing, and the demonstrated integration with existing developer tooling all point to a release worth watching, but the concrete license terms, hardware requirements, and real-world performance will only become clear once Qwen3.8-Max is in users' hands next week.