Image: METAL
Summary
- Alibaba officially unveiled Qwen3.8-Max on August 3. It's a 2.4-trillion-parameter MoE model with a 1-million-token context window, but the number of parameters actually activated during inference was not disclosed.
- When METAL LAB asked the model directly on Qwen Studio, it replied that it "cannot say for certain." Foreign outlet estimates are also split, ranging from roughly 95 billion to roughly 22 billion.
- Only the 27B model has open weights. Every flagship since Qwen3-Max in September 2025 has been API-only without exception, and 3.8-Max follows the same pattern.
Alibaba officially unveiled its flagship model Qwen3.8-Max on August 3. It's built on a mixture-of-experts (MoE) architecture with a total of 2.4 trillion parameters and a 1-million-token context window. This comes just two weeks after the model first appeared as a preview on July 19.
The numbers are big. But there's one number this announcement never confirms.
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 during a single inference pass. We added a caveat: if it wasn't certain, it should say so. Here's the first line it returned.
I cannot say for certain the total number of parameters in Qwen3.8-Max or the number of parameters that actually activate during a single inference pass.
The explanation that followed was, if anything, quite accurate. It defined 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 specific input token." It also noted that memory-loading cost scales with total parameters, while per-token compute cost scales with active parameters.
The Korean-language quality was good — 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 own spec sheet would be part of its training data. The real issue is that Alibaba itself never disclosed this number either.
As a result, foreign outlet estimates diverge. InfoWorld reported that roughly 95 billion parameters activate during inference, while strategy analysis outlet FourWeekMBA put the figure at around 22 billion per token. That's a gap of more than fourfold.
Why does this number matter in an MoE model? The 2.4 trillion figure describes how large the model was built; the active parameter count describes how much it costs to run every single time. When enterprises weigh adoption, it's the latter number they actually calculate against. When only the headline figure is large and the real operating figure stays undisclosed, the announcement leans closer to a scale flex than a genuine performance disclosure.
Qwen's three-year trajectory — from open to closed
Qwen built its reputation on being open-weight. Starting with Qwen-7B in August 2023, Alibaba successively released 14B and 72B versions, then kept opening its flagship models through Qwen2, Qwen2.5 in 2024, and Qwen3 in 2025. It was close to a symbol of China's open ecosystem.
| Date | 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 vertically and a clear line emerges. Since Qwen3-Max in September 2025, every Max-series model without exception has been API-only. What's been opened are the small-to-mid-size models — 27B, 397B. This time too, what will be 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 announcement, but what it actually refers to isn't the model in the headline.
Benchmarks came out mixed, category by category
Alibaba set Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Thinking as its comparison points. The results didn't lean one way.
| 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 came out ahead on terminal 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 coming "right after Fable 5." Worth noting that this wasn't an announcement claiming the top spot.
The first flagship since the team lead's departure
There's context that can't be left out when reading this model. On March 5 of this year, Alibaba approved the resignation of Lin Junyang (林俊旸), the head of Qwen's technical team — the person most responsible for making Qwen what it is today.
A day later, Alibaba announced an AI model task force jointly led by CEO Wu Yongming (吴泳铭), Alibaba Cloud CTO Zhou Jingren (周靖人), and Wu Zeming. In an internal letter, Wu Yongming said the company would pool group resources to accelerate foundation model development, and stated flatly that it would "continue to maintain its open-source model strategy." The move had a clear damage-control flavor. Post-training work was subsequently taken over by Zhou Hao, a former DeepMind researcher.
Qwen3.8-Max is the first flagship to emerge from that new structure. And the promise made five months ago has now been carried out in a form that opens the 27B and closes off the 2.4 trillion.
Next week is when the real announcement happens
Forrester's Charlie Dai assessed that enterprises now have a credible alternative to proprietary frontier models, particularly in software engineering. Gartner's Nitish Tyagi said the combination of open weights, MoE architecture, and a 1-million-token context window marks a step toward making AI-assisted development economically viable.
Still, the point analysts converged on is close to the real summary of this announcement: opening an API and opening weights are two separate acts.
What gets released next week, under what license, and at what scale, will determine how much weight this announcement actually carries. Until then, the 2.4 trillion figure remains a number that can't be verified.
Comments