AI & Services

2.4 Trillion Parameters Going Open: Why Qwen's Bold Move Threatens OpenAI.

Alibaba will release open weights for Qwen3.8-Max (2.4T) on August 12. It is the first open release of a Max-class model and a new benchmark for the open-weights camp.

· Huuush

2.4 Trillion Parameters Going Open: Why Qwen's Bold Move Threatens OpenAI

Key takeaways

  1. 1Alibaba is releasing Qwen3.8-Max, with 2.4 trillion total parameters, as open weights. It is the first time a "Max-class" model has been released with open weights.
  2. 2The release is scheduled for 10 a.m. (UTC+8) on August 12, 2026, alongside a 27B dense model for local deployment. The 2.4T model itself needs a data-center-class environment to run.
  3. 3API pricing is low at $2 input and $6 output, and multimodal and document work are strengths, but coding benchmarks are relatively weak and the license is not yet confirmed.

Alibaba's Qwen3.8-Max will be released as open weights. With 2.4T (2.4 trillion) total parameters, it is the first time a "Max-class" model has gone open weights. The release is scheduled for 10 a.m. on August 12, 2026 (UTC+8, or 02:00 UTC).

Qwen3.8-Max model architecture diagram
Qwen3.8-Max: 2.4T parameters in an MoE design, with about 95 billion active parameters per token (source: X post)

Unpacking 2.4 Trillion Parameters: The Specs First

  • Total parameters: 2.4T (sparse MoE, about 95B active parameters per token)
  • Context window: up to 1 million tokens
  • Multimodal: text, image, and video input
  • Where: Hugging Face and ModelScope, Aug 12 at 10 a.m. (UTC+8)
  • Also released: Qwen3.8-27B (27B dense) for local deployment

Max-Class, Once Locked Behind an API, Is Now Downloadable

Until now, Qwen's Max-class models were available only through an API. An open-weights release means you can download and run a top-tier model yourself. However, 2.4T is a multi-node, data-center-class artifact, so most people cannot run it locally. The 27B model handles local use.

$2 In, $6 Out: What's Different at This Price

  • API price: $2 input / $6 output per 1 million tokens, about 60% to 80% cheaper than GPT-5.6 Sol
  • Strengths: multimodal and document work (PaperBench 93.0 vs. 88.8 for Claude Fable 5)
  • Weaknesses: coding benchmark SWE-bench Pro 67.7 (Fable 5 scores 80.0)
  • Autonomous coding: 265 commits, 127 PRs, and 151 issues in 16 days of unattended coding

From Price Wars to Real-World Productivity: The Open Camp's Next Move

Open-weights players such as Qwen and Kimi are expanding the debate from price and performance into real-world productivity. This Max release is the biggest single event in that trend. The license (whether it is Apache 2.0) is not yet confirmed, but earlier open Qwen releases were Apache 2.0.

After the 2.4T Release, Who Can Actually Run It?

The release is symbolic because the open-weights camp has cleared the wall around API-only top-tier models for the first time. But practical use of 2.4T is limited to data-center operators. What matters is checking the license and real-world performance after the August 12 release.

Related: Alibaba Cloud · Quartz · TechRepublic

Frequently asked questions

What kind of model is Qwen3.8-Max?
It is a sparse MoE model with 2.4 trillion total parameters and about 95B active parameters per token. It supports up to a 1 million-token context and text, image, and video input.
When and where will it be released?
It is scheduled for 10 a.m. (UTC+8, which is 02:00 UTC) on August 12, 2026, on Hugging Face and ModelScope.
Can I run it on a personal computer?
Not the 2.4T model. It is a multi-node, data-center-class artifact that most people cannot run locally. For local use, the 27B dense model released alongside it is the option.
What is the license?
Whether it will be Apache 2.0 has not been confirmed. Earlier open Qwen releases were Apache 2.0.

Related articles