Alibaba says its most capable model ever goes open-weights next week

Qwen3.8-Max would be the first Max-class flagship Alibaba has opened. The promise lands a week after Kimi K3's weights shipped, days after OpenAI slashed prices, and as Congress audits Chinese models inside US apps.

Agent-drafted, human-editedPublished August 3, 2026

Reviewed by Trevor Brown

Alibaba's Qwen team announced Qwen3.8-Max late Sunday, calling it "our most capable model to date" and promising that its open weights "will be released" next week, with a smaller Qwen3.8-27B going open alongside it.1

Qwen

@Alibaba_Qwen

·Follow

📢Meet Qwen3.8-Max — our most capable model to date. Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!🎉 Qwen3.8-Max, a new bar for coding and cowork at 2.4T parameters: - Autonomous coding: 10+ days of Show more

2:15 AM · Aug 3, 2026

20.4K

Read 1.0K replies

The Qwen team's August 2 announcement on X: open weights promised within the week for both Qwen3.8-Max and the smaller Qwen3.8-27B. The specifications and benchmark chart are Alibaba's own.

Why it matters: Alibaba has not previously published weights for a Max-class model; its release post calls this "the first time we will open-source the weights of a Qwen-Max-class model."2 None appears in its public catalog.3 The announcement is the third major Chinese open-weights move in a week, and the clearest sign yet that the open-weights race is being run on a Chinese release schedule while Washington argues about it.

What Alibaba claims: 2.4 trillion parameters and "a new bar for coding and cowork."1 The centerpiece is autonomy: the company says the model ran "10+ days of self-evolving development, from empty folder to production without hand-holding," and points to a public GitHub repository as the trace.1 The claims and the benchmark chart are Alibaba's own; the model is live on QwenCloud, but the weights are not out.2

The price is the tell: Alibaba put API pricing in the announcement itself, $2 per million input tokens and $6 per million output.1 That matches the input price OpenAI set for GPT-5.6 Terra , when it cut Terra 20 percent and its budget Luna tier 80 percent, and undercuts Terra's output price by half.

Provenance & disclosure

AI use

Agent-drafted, human-edited

Disclosure

Agent-drafted for review and editing by Trevor Brown. Still in the Loop read Alibaba's announcement directly on X and its release post on qwen.ai, and verified the quotes, the pricing, and the open-weights commitment against those originals. The cited GitHub repository, the House committee release, and the committees' letter to DoorDash were fetched and read in full, and an independent verification pass re-checked names, dates, quotes, and links against their sources. Model specifications and performance claims are Alibaba's own; the weights had not been published at the time of writing.

Reviewed by

Trevor Brown — Editor-in-chief

Sources

1.Social reactionThe announcement, the "our most capable model to date," "a new bar for coding and cowork," and "10+ days of self-evolving development" language, the 2.4-trillion-parameter figure, the next-week open-weights commitment for Qwen3.8-Max and Qwen3.8-27B, the cited trace repository, and the $2 / $6 per-million-token pricing, from the Qwen team's post, @Alibaba_Qwen, August 2.
2.Official sourceThe "first time we will open-source the weights of a Qwen-Max-class model" line and the model's availability through QwenCloud, from Alibaba's release post, Qwen3.8-Max on qwen.ai.