Alibaba Unveils Qwen3.8 Max, but How Many of Its 2.4T Parameters Actually Fire Stays Undisclosed

Alibaba officially unveiled its flagship model Qwen3.8-Max on August 3. It's a mixture-of-experts (MoE) architecture with 2.4 trillion total parameters and a 1-million-token context window. The launch comes just two weeks after a preview debut on July 19.

The numbers are big. But there's one number this announcement never nails down.

We Asked the Model Itself

METAL LAB logged into Qwen Studio (chat.qwen.ai) and asked Qwen3.8-Max directly how many total parameters it has and how many parameters activate per inference. We added a caveat: if it wasn't sure, it should say so. The first line that came back was this:

I don't know for certain the total number of parameters in Qwen3.8-Max or the number of parameters actually activated per inference call.

The explanation that followed was, if anything, precise. It described total parameters as "the sum of all experts and shared layers the model holds," and active parameters as "the scale of parameters actually used in computation through the subset of experts the router selects when processing a given input token." It went on to note that memory-loading cost scales with total parameters, while per-token compute cost scales with active parameters.

The Korean output was clean — no translation-ese, and terms like router, forward pass, and shared layer were used correctly in context. Note that when not logged in, a sign-up prompt appears after just one exchange.

Nobody Knows the Active Parameter Count

It's not strange for a model to be unaware of its own specs — there's no reason its training data would include its own spec sheet. The real issue is that Alibaba itself never disclosed this number.

As a result, outside estimates diverge. InfoWorld reported that around 95 billion parameters activate during inference, while strategy analysis outlet FourWeekMBA put it at roughly 22 billion per token. That's a gap of more than 4x.

Why does this number matter for an MoE model? The 2.4 trillion figure describes how big the model was built; the active parameter count describes how much it costs to run every single time. When enterprises run the numbers before adopting a model, it's the latter figure they're calculating with. When only the headline number is big and the actual operating figure stays undisclosed, the announcement looks less like a performance disclosure and more like a scale flex.

Three Years of Qwen — A Trajectory of Opening, Then Closing

Qwen built its name on open weights. Starting with Qwen-7B in August 2023, Alibaba successively released 14B and 72B versions, then kept its flagship models open through Qwen2, 2.5 in 2024 and Qwen3 in 2025. It was close to a symbol of China's open ecosystem.

Timing	Model	Weights
Aug 2023	Qwen-7B	Open
Sep 2024	Qwen2.5	Open
Apr 2025	Qwen3	Open
Sep 2025	Qwen3-Max	API-only
Feb 2026	Qwen3.5 (397B-A17B)	Open
Apr 2026	Qwen3.6-Plus / 3.6-Max Preview	API-only
Apr 2026	Qwen3.6-27B	Open
May–Jun 2026	Qwen3.7-Max / 3.7-Plus	API-only
Aug 2026	Qwen3.8-Max	API-only

Read the table top to bottom and a clear line emerges. Since Qwen3-Max in September 2025, every single Max-tier model has been API-only. What did open up were small-to-mid sized models like the 27B and 397B versions. This time, too, what gets released on Model Studio next week is Qwen3.8-27B — not the 2.4-trillion-parameter flagship itself.

The phrase "open weights" keeps getting attached to this launch, but what it actually points to isn't the model in the headline.

Benchmarks Were Mixed, Category by Category

Alibaba set up Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Thinking as comparison targets. The results didn't lean uniformly in one direction.

Benchmark	Qwen3.8-Max	Comparison
Terminal-Bench 2.1	86.6	Claude Opus 4.8 — 84.6
SWE-bench Pro	67.7	Claude Fable 5 — 80.0
PaperBench	93.0	—
GPQA Diamond	92.6	—

It led on terminal-based tasks. But on SWE-bench Pro, which requires fixing actual issues in real repositories, the gap widened to more than 12 points. Alibaba itself described its own model as ranking "just behind Fable 5." It's worth noting that this was not an announcement claiming the top spot.

The First Flagship Since the Team Lead's Departure

There's context here that can't be left out when reading this model. On March 5 of this year, Alibaba approved the resignation of Lin Junyang, the head of the Qwen technical team — the person who made Qwen what it is today.

A day later, Alibaba announced an AI model task force to be jointly led by CEO Wu Yongming, Alibaba Cloud CTO Zhou Jingren, and Wu Zeming. In an internal letter, Wu Yongming stated the company would pool group-wide resources to accelerate foundation model development, explicitly stating that "we will continue to maintain our open-source model strategy." The move had the distinct look of damage control. Post-training was handed to Zhou Hao, a former DeepMind researcher.

Qwen3.8 Max is the first flagship to emerge from that new setup. And the promise made five months ago has, in practice, taken the shape of opening the 27B while keeping the 2.4-trillion-parameter model closed.

Next Week Is the Real Announcement

Forrester's Charlie Dai assessed that enterprises now have a credible alternative to proprietary frontier models, particularly in software engineering. Gartner's Nitish Tyagi noted that the combination of open weights, an MoE architecture, and a 1-million-token context window marks a step toward making AI-assisted development economically viable.

Still, a line analysts kept raising in common comes close to summarizing this entire launch: opening an API and releasing weights are two separate acts.

What opens next week, under what license, and at what scale, will determine how much weight this announcement actually carries. Until then, 2.4 trillion remains a number that can't be verified.