Alibaba has released Qwen3.8-Max, the most capable artificial intelligence model it has ever built, marking the first time the company has made a Max-class Qwen model available as open weights.

The model weights will be published on Hugging Face and ModelScope the week following the Monday launch. With a total of 2.4 trillion parameters—95 billion of which are activated at any given time—the model uses a Mixture-of-Experts architecture that routes each computation through a subset of specialized subnetworks rather than engaging the full model at once. That design preserves overall capacity while reducing inference costs, and it is intended to allow smaller businesses and research labs with sufficient hardware to run a state-of-the-art system without incurring the per-token costs associated with calling a large-scale datacenter API.

Endurance-Focused Performance Claims

Rather than leading with conventional benchmark comparisons, Alibaba emphasized the model's sustained autonomous capabilities. According to the company, Qwen3.8-Max spent 16 days independently building a coding tool, producing 265 commits, 127 pull requests, and 151 issues without human intervention. It also reportedly reproduced a research paper for which it had no source code over five days, then exceeded the paper's original results by 2.7 points. In a 24-hour machine learning competition, the model outperformed 458 of 526 participating human teams.

Compatibility With Rival Coding Tools

Qwen3.8-Max ships with setup instructions for Anthropic's Claude Code and OpenAI's Codex—both competing platforms. Alibaba's API is compatible with both companies' protocols, and most of the model's coding benchmarks were conducted inside Claude Code.

The benchmark results, however, are mixed. Across 31 text-based tests on Alibaba's own evaluation table, Anthropic's Fable 5 secured first place in 15 categories, OpenAI's GPT-5.6 Sol took nine, and Qwen3.8-Max won seven. On 12 coding-specific tests, Qwen won exactly one.

Where the model stands out is cost efficiency. Alibaba states that Qwen3.8-Max operates at approximately 30% of what Anthropic charges for Claude Fable 5, meaning that even when additional iterations or reasoning steps are required, the total cost of completing a task remains significantly lower.

The rankings shift substantially on multimodal tasks involving documents, video, and spatial reasoning, where Qwen3.8-Max leads most of the benchmark categories.

Strategic Reversal on Open Access

The release represents a notable change in direction. In April, Alibaba eliminated the free tier of Qwen Code, with the team moving toward closed, paid models following leadership departures. A Decrypt review of Qwen 3.7 Max previously noted that while the Plus version would be open, the Max variant was expected to remain accessible only through the API.

The timing of the open-weight release aligns with broader shifts in the AI landscape. Chinese open-weight models grew from under 2% of tokens processed on OpenRouter in late 2024 to approximately 61% by mid-2026. Qwen has surpassed Meta's Llama as the most self-hosted model globally.

Open-weight availability means organizations can download, audit, and run the model on their own infrastructure—an appealing proposition for enterprises with data-sovereignty requirements or those seeking to avoid vendor lock-in, particularly as geopolitical tensions complicate cross-border API dependencies.

Meanwhile, Washington restricted Fable 5 and Mythos 5 under export controls in June, and Beijing is reportedly considering its own limitations on Chinese models being deployed overseas.

While Qwen3.8-Max trails competitors on several benchmark leaderboards, Alibaba appears to be prioritizing distribution and accessibility—positioning the model as a free, downloadable alternative that approaches the performance of leading proprietary systems at a fraction of the cost.