AI News
Alibaba launches Qwen4-Max properly: $1.50 per million tokens, 1M context, and no open weights yet
Alibaba's Qwen4-Max is generally available from November 16, five days after a short preview that ran on a subscription plan only. The model, listed on Alibaba Cloud Model Studio as qwen4-max, is the first of Alibaba's new Qwen4 family and the successor to Qwen3.8-Max. Anyone can now call it through the public API, with published per-token prices, in six regions.
What changed from the preview
The preview, which opened on November 11, was narrow: Token Plan subscription only, Singapore only, no pay-as-you-go, and no published per-token price. It runs until midnight on November 25 (UTC+8) and is then retired. The generally available model runs in Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong, with public pricing. Two details changed with it: the maximum output is now 262,144 tokens for thinking plus answer combined (131,072 without thinking), down from the preview's 393,216, and the default reasoning effort is high rather than the preview's xhigh.
The price
In Singapore, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens, flat across the full 1M-token window, with thinking and non-thinking charged the same. The other regions — Beijing, Hong Kong, Frankfurt, Virginia and Tokyo — are priced slightly lower, at $1.413 and $4.24. That is 25% below Qwen3.8-Max, which costs $2.00 and $6.00. Singapore also carries a 1M-token free quota, valid for 90 days from activation. Cached input is cheaper still: implicit cache input costs $0.075 per million tokens.
What the model is
Qwen4-Max is a sparse mixture-of-experts model: a 2.4-trillion-parameter backbone with 71 billion parameters activated per token, plus 328 billion N-gram embedding parameters. Alibaba says that is 25% fewer active parameters than Qwen3.8-Max at the same scale. It takes text, image and video as input and returns text, with thinking on by default at low, high or xhigh effort. It supports tool calling and parallel tool calls, web search, structured output and caching, and works through OpenAI, Anthropic and DashScope-compatible APIs. Batch inference, fine-tuning and custom model deployment are not supported. The model is also available through Qwen Chat, Qwen Studio, QwenWork and Qoder, and on OpenRouter and the Vercel AI Gateway from day one.
The benchmarks, honestly
Alibaba's own numbers put Qwen4-Max ahead of its predecessor and, on some tests, ahead of rivals. On Alibaba's in-house Work360 benchmark, covering 721 professional fields, it scored 52.69 against 47.74 for Qwen3.8-Max, and Alibaba says the average cost per task fell from $4.43 to $1.67. On DeepSWE it scored 75.7, ahead of Claude Opus 5.5 at 74.2. But the same launch post shows it trailing on other tests: 48.5 against 59.6 for Opus 5.5 on Terminal-Bench 4.0, and 31.5 against 65.0 on ProgramBench. The first independent scores, published on launch day, tell a similar mixed story: a preliminary 1,750 points on LMArena's Code Arena WebDev board, good for sixth place, and 43 on the Artificial Analysis Intelligence Index against 59 for Opus 5.5. These are early scores and will move.
Safety, and what stays closed
Alibaba gave safety unusual prominence in the launch. Reuters reports the training included security checks alongside coding tests, cyber training data was filtered by intent — keeping vulnerability detection and fixes, removing material that strengthens attacks — and an internal review found no capabilities posing large-scale cybersecurity risks. On the Last Secure Code Bench, the secure-code pass rate rose to 31.8% from 22.0% for Qwen3.8-Max. What Alibaba did not open is the Max model itself: there are no downloadable weights for Qwen4-Max, and the company has said nothing about releasing them. Smaller Qwen4 models will get open weights, the announcement says, but it gives no date, no sizes and no licence.
What to watch
Two things matter next. First, whether the promised smaller open-weight Qwen4 models arrive, and under what licence — that decides whether Qwen4 reaches the home and small-business users who ran Qwen3.8 models locally. Second, how the independent benchmarks settle once more than launch-day scores exist. For now, Qwen4-Max is a cheaper, faster successor to Qwen3.8-Max that you can call from six regions — and a model whose best numbers are still Alibaba's own.