AI Knowledge HubQwen 4 Max Class Analysis: How Large a Flagship Will Follow the 2.4T Predecessor?
Qwen 4 Max Class Analysis: How Large a Flagship Will Follow the 2.4T Predecessor?
With the official launch of Qwen 4 Max imminent, we gauge the class of the upcoming flagship through the specs of its predecessor, Qwen 3.8 Max. This analysis is based on the roadmap revealed at the Apsara Conference 2026 and on actual data verified on GitHub/Hugging Face.
1. Predecessor (Qwen 3.8 Max) Spec Analysis
These are the specs of the flagship Qwen 3.8 Max currently in service on the market.
| Item | Spec |
|---|---|
| Total parameters | 2.4T (2.4 trillion) |
| Architecture | Sparse Mixture-of-Experts (MoE) |
| Active parameters | 95B per token |
| Context window | 1M tokens (up to 983,616 tokens) |
| Max output | 131,072 tokens |
| Input cost | $2.00 / 1M tokens |
| Output cost | $6.00 / 1M tokens |
| Multimodal | Text, image, and video input support |
| License | MIT-equivalent (open weights) |
Sources: Alibaba Cloud Model Studio, GitHub (AlibabaCloud-Official/Qwen3.8-max), MarkTechPost, DevelopersDigest (2026-08-03)
2. Where Qwen 4 Is Headed: The Apsara Conference 2026 Roadmap
This is what Alibaba CEO Eddie Wu announced at the Apsara Conference in Hangzhou on September 22, 2026.
Qwen 4 Is Currently in Training
According to Alibaba's official announcement, Qwen 4 is "in training," and the release schedule, model size, and license terms have not yet been disclosed.
Qwen 4.5 and Qwen 5 Series: 5-10T Parameters
Eddie Wu stated that the plan is "to train new models in the 5-10 trillion parameter range," and that "the goal is to complete more complex tasks with longer time horizons and to advance toward ASI." That is 2-4 times the scale of Qwen 3.8 Max (2.4T).
Sources: NYU Shanghai RITS, Reuters, Alizila (2026-09-22)
Qwen3.8-Flash-Next: A Preview of the Qwen 4 Architecture
There is a model that revealed its architecture ahead of the official launch.
| Item | Spec |
|---|---|
| Main model | 125B MoE |
| N-gram embedding | 51B |
| Active parameters | 6B per token |
| Release date | 2026-08-26 |
| License | Open weights |
Key architectural changes:
- Gated DeltaNet + QSA hybrid attention: efficiently compresses history and dramatically cuts attention cost on long sequences
- Gated Residual: expands residual connections into four branches and controls them with dynamic gates
- N-gram Embedding: uses local context to expand model capacity with almost no extra computation
- Muon optimizer: improves the division of roles between accuracy and AdamW
Sources: GitHub QwenLM/Qwen3.8-Flash-Next, Hugging Face, inferenceX SemiAnalysis
3. Hardware Infrastructure: V900
A next-generation AI accelerator developed by Alibaba T-Head.
| Item | V900 | M890 (predecessor) |
|---|---|---|
| Compute performance | 3x | 1x (baseline) |
| Memory | 216 GB HBM | 144 GB HBM |
| Inter-chip bandwidth | 1,200 GB/s | 800 GB/s |
| Large-scale cluster | Supports up to 500,000 units | - |
| Commercialization | Mass production scheduled for Q1 2027 | - |
Eddie Wu stated that "the V900 is currently the most powerful AI chip in China." The M890 supernode is already handling inference for foundation models of more than 2 trillion parameters, and commercial deployment begins this quarter.
Alibaba Cloud's global data center capacity is expected to surpass 20GW by 2032 (10 times that of 2022).
4. State of the Open-Source Ecosystem
| Metric | Figure |
|---|---|
| Cumulative open-source models | More than 460 |
| Cumulative downloads | Surpassed 3 billion |
| Derived models | More than 300,000 |
Alibaba has so far been the most aggressive in open-sourcing Qwen-related models through Hugging Face and elsewhere. Eddie Wu noted that "the open-source Qwen 27B has become the most popular model among developers worldwide."
5. Verified Benchmark Figures
Since the Qwen 4 generation has not yet been officially released, this lays out the actual benchmarks of the currently published Qwen 3.x series.
| Model | SWE-bench Verified | Other key benchmarks |
|---|---|---|
| Qwen3.8-Max (2.4T) | - | Text Arena 5th, Vision Arena 2nd |
| Qwen3.8-27B | - | SWE-bench Pro 61.7, IFBench 79.5, OSWorld 84.3 |
| Qwen3.6-27B (Dense) | 77.2% | Terminal-Bench 2.0 59.3% (Claude Opus 4.6 level) |
| Qwen3-Coder-Next (80B-A3B) | 70.6% | - |
| KAT-Coder-V2.5 (35B-A3B) | 69.4% | - |
Sources: LLMCheck, regolo.ai, tokenmix.ai, public benchmark data
6. Limitations to Watch (Community-Verified)
Behind the powerful performance there are limitations unique to the Qwen series.
Strong stubbornness (ego): When it outputs incorrect information, it has a strong tendency to refuse to admit it and to insist to the end, making system prompt tuning essential.
Strict self-censorship: As is typical of Chinese models, a censorship system triggers and immediately halts the conversation when politically or socially sensitive issues are included in the prompt. It is known to have tighter censorship standards than DeepSeek.
7. Criteria for Telling Fake Rumors Apart
A rumor is currently circulating that "Qwen 4 Coder 32B achieved 82% on SWE-bench Verified," but this has been confirmed to be a fake spec.
Grounds for Determining It Is Fake
- Not in official channels: It has never once been announced on Alibaba's blog, GitHub, or the Hugging Face Organization
- No weight files: Every actual Qwen release publishes downloadable weights on Hugging Face without exception
- Unrealistic benchmark: If a 32B model achieved 82%, it would break the record of a 276B model (Inkling-Small, 80.2%), which is currently technically impossible
- Stolen specs: The combination of ~32B MoE + ~3B active + Apache 2.0 is Qwen3.6-35B-A3B with only the version number changed
LLMCheck officially announced that it "removed this model from its records."
Source: LLMCheck "Qwen 4 Coder: Why You Can't Download It" (revised 2026-08)
8. Summary: What Class Is Qwen 4 Max?
| Category | Qwen 3.8 Max | Qwen 4 Max (estimate) | Qwen 4.5/5 (planned) |
|---|---|---|---|
| Total parameters | 2.4T | Undisclosed (estimated 3T+) | 5-10T |
| Architecture | MoE | GDN+QSA hybrid | Undisclosed |
| Active parameters | 95B | Undisclosed | Undisclosed |
| Open weights | Yes | Undisclosed | Undisclosed |
The exact specs of Qwen 4 Max must wait until launch. However, considering that a 2-4x expansion over the predecessor's 2.4T is planned and that the V900 (a 500,000-unit cluster) is commercialized around the same time, Qwen 4 Max is highly likely to be an ultra-large class in the 3T-5T range.
Whether Alibaba will keep open weights or change its license policy is still unconfirmed. Up to Qwen 3.8 Max it maintained open source under an MIT-equivalent license, so whether the same stance continues with Qwen 4 is the key point to watch.
Comments (1)
To start from the conclusion, this piece estimates the Qwen 4 Max size at 3T-5T based on Qwen 3.8 Max's 2.4T spec, and organizes its roadmap and the criteria for spotting fake rumors with a verification-centered approach. Section 7, which refutes the "Qwen 4 Coder 32B 82%" rumor on the grounds of missing weights and unrealistic benchmarks, is especially useful. That said, "μ 무" in lines 64 and 76's "μ 무 V900" means "none at all" in Korean and may be mistaken for a chip name, so the transliteration should be fixed, and line 26's "κ·Όμ MIT" (close to MIT) reads better as "MITμ μ€νλ" (MIT-equivalent).