Qwen4-Max Day-One Launch: 1M Context Is Just the Entry Ticket—Can It Actually Finish Long Tasks?
Qwen4-Max is available on day one via Novita AI. 2.4T MoE, 977K context, native vision, 128K output, $1.413/M input tokens, $4.24/M output tokens. The real test is whether it can reliably deliver on long-horizon tasks and tool calls.
Qwen4-Max launched on Novita AI today and is available from day one. Officially, it is a MoE flagship with 2.4T total parameters and 71B activated; the page lists a 977K-token context, while the broader messaging says 1M; maximum output is 128K tokens.
It supports native vision, with inputs covering text, images, and video, and text output; reasoning, structured outputs, and function calling are all on the supported list. The API is serverless and billed per token: input $1.413/M tokens, output $4.24/M tokens, cache reads $0.071/M tokens, 5-minute cache writes $1.767/M tokens, and 1-hour cache writes $2.827/M tokens. It is OpenAI-compatible and also supports the Anthropic API.
A day-one launch is no big deal. Whether it can finish long tasks is the real battle. The official pitch centers on complex coding, professional tools, and long-horizon agent workflows, claiming the model can autonomously drive long-cycle projects and deliver complete output over dozens of days, covering professional fields such as law, finance, design, and scientific research. That description sounds more like selling an execution loop than selling chat.
In the replies to the tweet, some users asked what workload to run first with 1M context. Others pressed for details on 27B. Another user asked directly whether releasing a minor version every month might push researchers to overreach. Beyond the congratulations, these questions are closer to real-world use.
If you want to try it, the model ID is qwen/qwen4-max, the base URL is https://api.novita.ai/openai, and the Anthropic-compatible endpoint is https://api.novita.ai/anthropic. It is suitable for small-scale validation first with long-context codebase analysis, very long documents, long-video understanding, and long-chain tool calls. Context numbers are easy to spread, but bills and task completion rates are more honest.
Original post: https://x.com/Alibaba_Qwen/status/2122286761883959700
发布时间: 2026-11-16 19:27