Wink Pings

Qwen4-Max Launch: 2.4T MoE, 71B Activated Parameters, 1M Context Window, $1.5 per Million Input Tokens

Alibaba releases Qwen4-Max, featuring 2.4 trillion total parameters, 71 billion activated parameters, native vision support, and a 1M-token context window. Its input is priced at $1.5 per million tokens, cutting single-task costs on Work360 by 62%. We break down its core metrics and real-world implications separately below.

Alibaba has officially released Qwen4-Max. According to MarkTechPost's report, this is a closed-source model built with a MoE architecture, with 2.4T total parameters and only 71B parameters activated per single inference pass. It natively supports vision inputs and comes with a 1M-token context window. The input pricing is set at $1.5 per million tokens. The report also mentions that it reduces single-task costs on the Work360 platform by 62%.

Let's first look at several key figures. 2.4T refers to the total parameter count, while 71B refers to the activated parameters. MoE models distribute tokens to only a subset of expert modules, so the inference computing requirement is much closer to the activated parameter level, though storage and deployment specifications still depend on total parameters. The direct takeaway for developers is that this is an extremely large model, but you do not need to run all 2.4T parameters for every single call. The 1M context window means far longer documents, full codebases, or extended multi-turn task records can be included in one single request. Native vision support eliminates the extra work of assembling a separate external vision encoder, shortening the entire engineering pipeline.

On the pricing front, the public material only lists input tokens at $1.5 per million. This price point is well worth attention for a closed-source model with 2.4T total parameters. What is even more notable is the 62% single-task cost reduction on Work360: single-task cost is far closer to the actual budget concern of purchasers than per-token pricing, as it accounts for total spending to complete a specific task, required retries, and integrated tool call overhead.

Qwen4-Max is a closed-source model, which is fundamentally different from open-source models that allow full weight download. It is suitable for teams that do not want to maintain their own large model inference clusters, but need long context and vision capabilities to carry out initial verification. The remaining metrics to track include output token pricing, rate limiting rules, tool call stability, and whether the full 1M-token context window can be reliably used to its full extent in real-world tasks. Parameter scale still has news value, but actual cost and output quality are the core factors that determine whether this model will remain in production environments long-term.

Source: MarkTechPost, November 16, 2026.

发布时间: 2026-11-16 17:44