On November 16, 2026, Alibaba released Qwen4-Max, the first flagship model based on the new Qwen4 architecture. The model relies on a MoE backbone with 2.4 trillion parameters, accepts text, images, and video, and handles contexts of up to 1 million tokens. According to Qwen, the average cost per task on the internal Work360 benchmark decreased by 62% while quality improved, and the model is already available via the QwenCloud API, Qwen Studio, QwenWork, and Qoder.


What Happened
Alibaba introduced Qwen4-Max, the first flagship on the Qwen4 architecture. The model's foundation is a MoE backbone with 2.4 trillion parameters, of which 71 billion are activated per token; additionally, the architecture utilizes 328 billion N-gram Embedding parameters. The context window is 1 million tokens, and it accepts text, images, and video as input. Three new components appeared in the attention block: Gated DeltaNet (GDN), Qwen Sparse Attention (QSA), and Gated Residual (GR). To demonstrate progress, Qwen uses the internal Work360 benchmark, covering 26 domains and 721 directions: Qwen4-Max scored 52.69 points compared to 47.74 for Qwen3.8-Max. According to the same vendor data, the average cost per task dropped from $4.43 to $1.67, a 62% reduction, while token consumption decreased by approximately 24–35% on Work360 and 48–49% on Terminal-Bench 4.0. API prices are set at $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 per million cache-read tokens. The model is already available via QwenCloud, Qwen Studio, QwenWork, and Qoder; the weights for Qwen4-Max itself are closed, but Alibaba promises to open up smaller models in the Qwen4 lineup later.
Context
The release can be conveniently read through two market trends it serves. The first is the economics of Mixture-of-Experts architectures: the total number of parameters is increased for quality, but only the active part is paid for, so the inference bill is determined by the 71 billion working parameters per token, not the full 2.4 trillion. The second is the high cost of long context with classical attention: this is what the new Gated DeltaNet, Qwen Sparse Attention, and Gated Residual components are aimed at, which, according to Qwen's description, reduce the cost of working with large context windows. Against this backdrop, the framework for comparing models is also changing: instead of price per token, the vendor promotes cost per task, because in multi-step agentic scenarios, the final bill depends on token consumption and the number of attempts, not just the rate. The scheme of "closed flagship now, open smaller models later" appears to be a deliberate product strategy: Max is immediately monetized via API, while open weights become a tool for ecosystem expansion.
Why This Matters for the Industry
The main signal for the industry is a new economics based on cost per task, rather than nominal parameters. If the claimed drop in average cost per task from $4.43 to $1.67 is confirmed, multi-step agentic scenarios—programming, document analysis, industrial workflows—will become profitable for a significantly wider range of companies, including those for whom such pipelines did not work out unit economics a month ago. Prices of $1.50 per million input tokens and $4.50 per million output tokens, with cache reading at $0.075, intensify price pressure on competitors in the flagship model segment: they will have to respond with either a comparable reduction in cost per task or price cuts. If the reduction in token consumption is confirmed by independent measurements, the combination of MoE, sparse attention, and N-gram Embedding could become a template for long-context flagships, and "cost per task" could become the main budgeting metric instead of "price per token." The promise to open smaller Qwen4 models creates a path for builders to "prototype on Max—production on open weights" with on-prem options and private data, and simultaneously provides the first opportunity for independent verification of the architecture outside vendor marketing.
Why This Matters for Users
The model is already available, so verification does not require waiting: Qwen4-Max works in QwenCloud API, Qwen Studio, QwenWork, and Qoder. A practical first step is to run 20–50 real tasks of your product in parallel on Qwen4-Max and Qwen3.8-Max and calculate the actual cost and token consumption per task, rather than an abstract rate: this is how the main release thesis about cost reduction is verified. It is worth starting the recalculation with scenarios where long context was the main blocker for economics—processing large documents and multi-step agentic chains; separately, it makes sense to measure the effect of cache reading at $0.075 per million tokens. For those for whom closed weights are not suitable, it is worth watching for the announcement of open smaller Qwen4 models, which Alibaba promises to release later.
What Is Still Unknown / Limitations
Key release figures are vendor self-reports. Work360 is an internal Qwen development: the comparison was conducted only with their own Qwen3.8-Max, and external control, a public task set, and protection against data contamination in sources are not described, so independent reproducibility is currently zero. The claim of a 62% reduction in cost per task is a composite metric: it mixes the reduction in API prices, the decrease in token consumption, and quality changes, and the decomposition of these factors has not been published. There are no ablations of GDN, QSA, and GR in the materials, so the contribution of each attention component cannot be separated from other changes—data and post-training. For production solutions, operational data is lacking: sources do not include latency (TTFT and generation speed), information on rate limits, SLA, and available regions—all of this will have to be determined independently during testing.
Sources
- Qwen4-Max: Your New Teammate Has Arrived (Qwen Team Blog)
- Alibaba Releases Qwen4-Max: 2.4 Trillion Parameters, Per-Task Cost Slashed by 60% — BigGo Finance
- Qwen4-Max benchmarks, pricing and specs: How it compares — kingy.ai
Author
Look at AI, editorial team
