Alibaba has launched Qwen4-Max, the initial model of its Qwen4 generation. It is a proprietary sparse mixture-of-experts system with a 2.4 trillion parameter backbone and 71 billion active parameters per token. The model handles text, images, and video within a 1 million token context window. Alibaba also reduced API pricing, reporting a 62% drop in average cost per task.
In this article
Key specifications
- Size: 2.4 trillion backbone parameters, 71 billion active per token, plus 328 billion N-gram embedding parameters. Context is 1 million tokens; maximum output is 393,000 tokens.
- Availability: Hosted only via QwenCloud API, Token Plan, Qwen Studio, QwenWork, and Qoder. There are no open weights, meaning self-hosting or quantized builds are not possible.
- Performance: A significant improvement over Qwen3.8-Max. It competes with frontier models on agentic tasks but trails in hard coding and vision.
- Top score: 21.3 on HLAB-Vals-1.1 (legal agent work), compared to 11.7 for Kimi K3.1, the next-best model.
What is Qwen4-Max?
Qwen4-Max is a multimodal mixture-of-experts language model from Alibaba designed for long-horizon professional work. Access is currently limited to a Token Plan-only preview available for one month.
How does the Qwen4-Max architecture work?
Qwen4 improves the sparse mixture-of-experts design in three areas: attention, residual connections, and embeddings.
- GDN + Qwen Sparse Attention (QSA): Each QSA layer pairs with three Gated DeltaNet layers. These compress history into fixed-size states. QSA then uses a lightweight indexer to select relevant context at micro-block granularity. This lowers attention and indexing costs on long sequences.
- Gated Residual (GR): Four parallel residual branches use dynamic read and write gates. The aim is stronger information flow and better training stability.
- N-gram Embedding: 328 billion extra parameters provide lookups of local token patterns at layer 2. This adds capacity with little extra compute.
Alibaba reports no loss spikes across the 2.4 trillion-scale pre-training run. On in-house tests, Qwen4-Max-Base shows 5 to 13% lower bits per byte than Qwen3.8-Max-Base. It achieves this while activating 25% fewer parameters.
How does Qwen4-Max perform on benchmarks?
Alibaba’s comparison chart covers 16 benchmarks. Rivals include Qwen3.8-Max-0902, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra, and Kimi K3.1. Several scores are cited from third-party leaderboards. None were independently reproduced at launch.
Qwen4-Max ranks first on 4 of 16 benchmarks:
- HLAB-Vals-1.1 (legal agents): 21.3, versus 11.7 for Kimi K3.1 and 9.2 for Opus 5.5.
- Agents’ Last Exam v1.1: 46.7, versus 43.4 for Opus 5.5 and Kimi K3.1.
- BioMysteryBench: 85.0, versus 84.4 for GPT-6.1 Sol.
- DeepSWE v1.1: 75.7, versus 75.5 for Kimi K3.1.
Claude Opus 5.5 leads 7 of the 16 charts. Qwen4-Max trails most clearly on coding and vision:
- ProgramBench: 31.5 versus 65.0 for Opus 5.5.
- Terminal-Bench 4.0: 48.5 versus 67.2 for Kimi K3.1.
- Chartography: 54.2 versus 71.1 for GPT-6 Astra.
- Humanity’s Last Exam: 51.7 versus 59.8 for Opus 5.5.
On Alibaba’s own Work360 benchmark, Qwen4-Max scores 52.69, up from 47.74 for Qwen3.8-Max. Work360 spans 26 top-level domains and 721 specialized fields.
Why is Qwen4-Max cheaper to run?
Two things stack: fewer tokens and lower prices. On Work360, token use drops about 24% at Xhigh effort and 35% at High effort. On Terminal-Bench 4.0, the drops are about 48% and 49%. Average Work360 cost per task falls from $4.43 to $1.67, a 62% cut.
List prices fell too. Input is $1.50 per million tokens, down from $2 for Qwen3.8-Max. Output is $4.50, down from $6. Implicit cache reads drop 70%, from $0.25 to $0.075.
Qwen4-Max vs competitors
| Metrics | Qwen4-Max | Qwen3.8-Max | Claude Opus 5.5 | GPT-6.1 Sol | Kimi K3.1 Pro |
|---|---|---|---|---|---|
| Total params | 2.4T backbone + 328B N-gram emb. | 2.4T | Not disclosed | Not disclosed | 2.8T |
| Active params | 71B | Not disclosed (blog implies about 95B) | Not disclosed | Not disclosed | 104B |
| Context | 1M | 1M | 1M | 1,050,000 | 1,048,576 |
| Max output | 393K | 131K | 128K | 128K | Not disclosed |
| Input modalities | Text, image, video | Not disclosed | Text, image | Text, image | Text, image |
| Weights | Proprietary, API only | Not disclosed | Proprietary | Proprietary | Open (Kimi K3.1 Pro License) |
| Input / output price (per 1M) | $1.50 / $4.50 | $2 / $6 | $4 / $20 | $2 / $10 | $2 / $10 |
| Cached input (per 1M) | $0.075 | $0.25 | $0.20 | $0.10 | $0.20 |
| DeepSWE v1.1* | 75.7 | 69.3 | 74.2 | 75.2 | 75.5 |
| Terminal-Bench 4.0* | 48.5 | 38.9 | 59.6 | 56.1 | 67.2 |
| Humanity’s Last Exam* | 51.7 | 43.9 | 59.8 | 54.0 | 51.1 |
| ProgramBench* | 31.5 | 28.0 | 65.0 | 52.5 | 25.0 |
| HLAB-Vals-1.1* | 21.3 | 15.4 | 9.2 | 6.3 | 11.7 |
*Scores from Alibaba’s Qwen4-Max chart. Qwen3.8 column is Qwen3.8-Max-0902. The chart labels Kimi as “Kimi K3.1”. Moonshot’s own card reports Terminal-Bench 4.0 of 65.7 for Kimi K3.1 Pro and 66.4 for Opus 5.5.
How can developers use Qwen4-Max?
The model ID is qwen

