Alibaba on Monday published benchmark scores it had withheld on July 19 for Qwen3.8-Max, its 2.4 trillion-parameter model. The launch post says Qwen3.8-Max has 95 billion active parameters. Many outlets have described it as a sparse mixture-of-experts design, which routes each query to a fraction of its parameters so serving it costs far less than its total size implies. Alibaba also set next week as the release window for the model weights.

What Changed

AI-generated summary, reviewed by an editor. More on our AI guidelines.

What Alibaba disclosed

On July 19, Alibaba called Qwen3.8-Max “second only to Fable 5,” according to an Implicator report based on South China Morning Post’s preview coverage. The report identified three missing disclosures: the activated-parameter count, the license and the weights date. Monday’s release supplied the count and date. Alibaba has not stated which license will govern the weights it plans to publish on Hugging Face and ModelScope.

The model has 2.4 trillion total parameters, about seven times as many as Qwen3.5, which Alibaba released in February. Its context window holds one million tokens, or roughly 750,000 words per query, according to the South China Morning Post. Alibaba said it will also release a smaller Qwen3.8-27B next week. Qwen3.8-Max will be the first Max-class Qwen model with downloadable weights.

The company benchmark table

Alibaba’s table put Qwen3.8-Max at 93.0 on PaperBench, the highest result among the models it listed. On Terminal Bench 2.1, Qwen scored 86.6, behind GPT-5.6 Sol at 88.8. The Decoder noted that these figures came from Alibaba’s internal runs and that independent verification is pending.

Crowdsourced Arena.AI results provide the only third-party placements available at publication. Qwen3.8-Max ranked fifth in Text Arena, second in Vision Arena and fourth in Frontend Code Arena. Alibaba selected those placements for its launch materials.

Price and investor response

Alibaba priced Qwen3.8-Max at $2 per million input tokens and $6 per million output tokens. Hong Kong-listed shares closed Monday up 7%, their largest gain in nearly a month, Bloomberg reported.

Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.

Vey-Sern Ling, managing director at Union Bancaire Privée, told Bloomberg, “Many investors continue to underestimate Chinese AI models because of US chip restrictions or general skepticism. In reality, the gap is probably much closer, and narrowing fast. Alibaba’s Qwen 3.8 is another proof point, following Kimi K3.”

Independent testing

The Kimi K3 model Ling cited was released by Moonshot with open weights on July 27. The model has 2.8 trillion parameters, but subsequent independent testing found that it fell well short of top Western models in cyber capabilities and complex math, according to The Decoder.

Andrew Yoon, a technical staff member at CivAI, told The Deep View, “Claims that recent Chinese models match or beat the US frontier are overstating cherry-picked benchmark results. This isn’t to say that the models are weak. Models like Qwen 3.8 Max, Kimi K3, and GLM 5.2 are all very capable, but they continue to lag the US frontier by a significant margin.”

Frequently Asked Questions

What did Alibaba actually disclose on Monday that it had not disclosed before?

On July 19 Alibaba called Qwen3.8-Max "second only to Fable 5" and published no benchmark scores. An Implicator report that day identified three missing disclosures: the activated-parameter count, the license, and the weights date. Monday's release supplied the activated-parameter count, 95 billion, and set next week as the weights window. The license remains unstated.

How large is Qwen3.8-Max, and how much of it runs at once?

The model has 2.4 trillion total parameters, about seven times as many as Qwen3.5, which Alibaba released in February. Forbes Middle East and TNGlobal described it as a sparse mixture-of-experts design, which routes each query to a fraction of those parameters. Alibaba's launch post puts the active count at 95 billion. Its context window holds one million tokens.

How did Qwen3.8-Max score on the benchmarks Alibaba published?

Alibaba's table put the model at 93.0 on PaperBench, the highest result among the models it listed, and 86.6 on Terminal Bench 2.1, behind GPT-5.6 Sol at 88.8. On crowdsourced Arena.AI leaderboards it ranked fifth in Text Arena, second in Vision Arena and fourth in Frontend Code Arena.

Can the benchmark numbers be trusted?

Not independently, not yet. The Decoder reported that the figures came from Alibaba's internal runs and that independent verification is pending. The Arena.AI placements are third-party, but Alibaba selected which ones to publicize. Andrew Yoon of CivAI told The Deep View that claims Chinese models match the US frontier are "overstating cherry-picked benchmark results."

What does Qwen3.8-Max cost, and how did the market react?

Alibaba priced the model at $2 per million input tokens and $6 per million output tokens. Hong Kong-listed shares closed Monday up 7%, their largest gain in nearly a month, Bloomberg reported. Jefferies analysts wrote that Alibaba's full stack of capabilities stands out.

AI-generated summary, reviewed by an editor. More on our AI guidelines.

Moonshot Releases Kimi K3 Weights as Amodei Rejects Open-Weight Ban
Moonshot AI released the weights for its Kimi K3 model on Monday, a system with 2.8 trillion parameters. Developers can now download, modify and self-host K3, which the report described as the world's
The Implicator
Zuckerberg Says US Should Not Ban Chinese AI Models, Warns of Regulatory Capture
Mark Zuckerberg told the Financial Times on Tuesday that the United States should not block Chinese AI models and warned that American frontier labs could gain too much influence over reviews of compe
The Implicator
Germany's Soofi S AI Model Tops All Open-Source Rivals on German Benchmarks
A German research consortium coordinated by the KI Bundesverband released Soofi S, an open-source German-English foundation model, this week, according to its pretraining report. In the team's tests,
The Implicator
AI News

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai