Alibaba Launches Qwen3.8-Max With 2.4 Trillion Parameters

The testable part of Qwen3.8-Max begins when its weights can actually be downloaded. If independent teams cannot reproduce the reported coding and long-horizon results with published settings over the next three to six months, or if deployment costs confine the model to a handful of cloud operators, the launch will demonstrate Alibaba’s ability to scale a model but not its practicality for enterprise-controlled infrastructure.

The Qwen team released Qwen3.8-Max on August 3, 2026. Alibaba describes it as a mixture-of-experts model with 2.4 trillion parameters in total and 95 billion active parameters per inference pass. It is available now through the QwenCloud API, while the team says the weights will follow next week. This would be the first open-weight release at the Qwen Max tier, allowing outside developers to inspect and deploy the model rather than testing only a hosted endpoint.

What 2.4 trillion means during inference

The total parameter count describes the model’s full capacity; the 95-billion active figure is closer to the portion used for each token. A mixture-of-experts design can therefore increase capacity without evaluating every parameter on every request. The published configuration also lists a one-million-token context window and low, medium, and xhigh reasoning-effort levels. Those settings affect memory use, latency, and cost, but the announcement does not disclose the required hardware, self-hosted throughput, training cost, or a complete per-token price comparison.

Alibaba positions Qwen3.8-Max for coding, cowork, research, and long-running agent tasks. Its table reports 86.6 on Terminal Bench 2.1, below GPT-5.6 Sol at 88.8 and above Claude Opus 4.8 at 84.6. It reports 93.0 on PaperBench. These are not uniform comparisons: Qwen’s Terminal Bench result uses Claude Code, an average over ten runs, a five-hour timeout, and 131,072 maximum tokens, while the other entries use their best published scores across different harnesses. The numbers show performance under named workflows, not a blind test under an identical compute budget.

Open weights will expose the deployment bill

The launch examples emphasize feedback loops in which the model proposes a method, writes code, runs an experiment, and changes the next attempt. In one research case, the loop ran for about 88 hours and produced a method that improved AIME24 by 2.7 percentage points over the reproduced paper method. Such an agent depends on more than model accuracy. Tool permissions, accumulated errors, GPU hours, and any unreported human intervention can materially change the cost and the final result. Alibaba does not publish a complete failure set or a full inference bill for the demonstrations.

Reuters independently reported that Alibaba had unveiled its most capable model to date and placed the launch alongside new Chinese models competing on both capability and price. That corroborates the event, while the detailed performance claims still originate mainly from Qwen’s own tests. The first checks are whether the weights arrive on schedule, what commercial rights the license grants, and how much GPU memory is needed to serve 2.4 trillion parameters. After that, third parties can rerun Terminal Bench 2.1 with the same harness, token ceiling, and time limit to see whether 86.6 survives outside Alibaba’s environment.

Sources:

• Qwen: Qwen3.8-Max: A New Bar for Coding and Cowork
• Reuters: Alibaba unveils its most capable AI model to date