Skip to the stories
Wednesday, 18 November 2026
Saved

Next edition

4:03

Headlines

Models & Products

Alibaba releases Qwen4-Max model with native vision MoE architecture

Alibaba Group Holding Ltd. has introduced Qwen4-Max, a 2.4-trillion-parameter sparse mixture-of-experts model featuring lower inference costs and upgraded attention mechanisms.

FTMQ SI, written by our newsroom0 views

Share
Picture: SiliconANGLE AI

Alibaba Group Holding Ltd. introduced Qwen4-Max, its latest proprietary large language model. SiliconANGLE reported that the release features an upgraded attention mechanism alongside architecture optimizations that improved training efficiency. MarkTechPost reported that Qwen4-Max is a 2.4-trillion-parameter sparse mixture-of-experts model with 71 billion active parameters, native vision capabilities, and a 1-million-token context window. [1][2]

MarkTechPost reported that input tokens for Qwen4-Max cost $1.50 per million, cutting the cost per task for Work360 by 62 percent. Wccftech reported that the model retains the same parameter count as its predecessor at a lower price point, utilizing a revamped native-vision mixture-of-experts architecture. However, Wccftech noted that Qwen4-Max failed to overtake Moonshot AI's Kimi K3.1 Pro. [2][3]

The Qwen4 design modifies the sparse mixture-of-experts stack by altering attention, residual connections, and embeddings. Each Qwen Sparse Attention layer operates alongside three Gated DeltaNet layers that compress context into fixed-size states. Four parallel residual branches use dynamic read and write gates, while n-gram embeddings provide token-pattern lookups at the second layer. Alibaba reported a stable pre-training run without loss spikes, recording 5 to 13 percent lower bits per byte than Qwen3-Max. [9]

Qwen is a family of language models developed by Alibaba Cloud. The New Stack reported that Alibaba did not release Qwen4-Max under an open-source license, keeping the model proprietary. [2][4][8]

Share

In short

  • Qwen4-Max contains 2.4 trillion total parameters and 71 billion active parameters.
  • The model supports native vision inputs and a 1-million-token context window.
  • Input pricing is set at $1.50 per million tokens.
  • Qwen4-Max uses a sparse mixture-of-experts architecture paired with Gated DeltaNet layers.

Sources

Every paragraph above points to the numbered items it rests on. Read the originals here.

  1. [1]Alibaba launches Qwen4-Max with training, attention optimizationsSiliconANGLE AI, 1d ago (the report this story comes from)
  2. [2]Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M ContextMarkTechPost, 1d ago
  3. [3]Alibaba’s New Qwen4-Max Fails To Overtake Moonshot’s Kimi K3.1 Pro, Retaining The Same Parameter Count With A Lower Price And Revamped Native-Vision MoE ArchitectureWccftech, 1d ago
  4. [4]With Qwen4-Max, Alibaba puts the open source jacket back on the hangerThe New Stack, 1d ago
  5. [5]Alibaba releases Qwen4-MaxConstellation Research, 1d ago
  6. [6]Alibaba challenges OpenAI and Anthropic with Qwen4-MaxTechzine Global, 1d ago

Background

  1. [7]DeepSeek (AI) on Wikipedia
  2. [8]Qwen on Wikipedia
  3. [9]Qwen4-Max on Grokipedia

Our newsroom writes these reports with the help of software, from the 9 sources listed and nothing else, and checks them against those sources. Facts can still be wrong or move on; the originals are the record. Spotted a mistake? Write to daniel@monsterkong.com.

Earlier reports of ours on the same people and subjects.

More in Models & Products

Get the day in one email

Reports like this one, the top news of the last 24 hours, every morning. Free, one email a day; readers can comment under every report.

By signing up you agree to our terms and privacy policy. Unsubscribe any time.

Comments

Loading

Join the conversation

Comments are open to readers of our daily email: the top news of the last 24 hours, every morning, free. Sign up and the comment box opens.

Already on the list? Enter the same address and we will send a sign-in link.

By signing up you agree to our terms and privacy policy. Unsubscribe any time.