Alibaba Launches Qwen3.8-Max, Its Largest AI Model Yet

Alibaba shipped Qwen3.8-Max at 2.4 trillion parameters. The headline is "largest yet." The actual story is that the open weights land next week.

The specs: 2.4T total MoE with 95B active, 1M token context, text plus image plus video. API at $2 per million input and $6 output, about 20 percent cheaper than 3.7. It sits at #4 on the Frontend Code Arena, roughly one point behind Claude Opus 5, and reportedly edges Fable 5 by about 2 percent on Terminal-Bench. Alibaba says it ran solo on a real project for 16 days and produced 265 commits and 127 PRs.

Three things stack here

1. The largest open model ever is about to exist

The current open-weights record is DeepSeek V4 Pro at 1.6T. This is 2.4T. And it is the first time Qwen has opened the Max tier at all. They used to open the mid-size models and keep the flagship behind an API. This is a strategy decision, not a technical one.

2. A receipt for a prediction from three days ago

In The Economist interview published July 29, Musk argued that Chinese labs got to the frontier on relatively little compute, so if they ever get a lot of compute they will probably lead. That was a deduction at the time. Reports now say Alibaba deployed this 2.4T model on entirely domestic infrastructure. I do not know whether his claim about lithography being closer than people think is correct. But a domestic stack carrying frontier-scale MoE in production moves that constraint down a notch.

3. The price war changed shape

Last week DeepSeek hit $0.14 per million input tokens. This week Qwen is $2 and $6. Those look opposed, but they are different attacks. DeepSeek is going after volume with a price nobody has to think about. Qwen is making the strongest tier self-hostable. For enterprises the second one may hit harder, because once flagship weights are downloadable the procurement question stops being "whose API do we buy" and becomes "which workloads are worth hosting ourselves." That is a completely different negotiating position.

What I take from it

As someone who watches token prices daily because we run a gateway:

Largest is not the point. Open is the point. When the top of the leaderboard is separated by one point, the assumption that you must be on a closed model starts to loosen. And this week Qwen, last week DeepSeek, the week before an 80 percent cut from OpenAI. No model stays best. The value was never in picking the right one. It is in having a layer underneath that trades off cost and capability for you.

---

Based on Alibaba's Qwen3.8-Max release and public reporting; benchmarks and pricing per official sources. Musk quote from The Economist's full-length interview, 29 July 2026.