Alibaba releases Qwen4-Max with 71B active parameters and lower prices
November 16: The Decoder adds a hybrid context design, an Intelligence Index of 43, a 46.7 Agents' Last Exam score, $0.075 cache hits, and the open-sourced Qwen-Live-Delivery-Plugin.
Alibaba's Qwen4-Max is out as a 2.4T MoE model with 71B active parameters, a 1M context, and cheaper API rates.
Alibaba has released Qwen4-Max, a sparse mixture-of-experts flagship for professional work with native vision and a one-million-token context. It uses a 2.4-trillion-parameter backbone, activates about 71 billion parameters per token rather than about 95 billion, and adds 328 billion N-gram embedding parameters. Alibaba says Work360 rose from 47.74 to 52.69, average task cost fell from $4.43 to $1.67, and list prices are $1.50 and $4.50 per million input and output tokens. The Decoder reports an Artificial Analysis Intelligence Index of 43 and 46.7 on Agents' Last Exam v1.1. Alibaba says filters keep defensive cybersecurity data and drop offensive attack material; analysts say buyers should verify the claims.
- High-effort mode used 35% fewer tokens; Work360 cost fell about 62%.
- Base model is about 5–13% better than Qwen3.8-Max-Base with fewer active parameters.
- Last Secure Code Bench pass rate rose from 22% to 31.8%, Alibaba says.
- Cache hits cost $0.075; smaller open-weight Qwen4 models are planned.
- Prompt retention is up to 30 days, longer for flagged non-China content.
- Alibaba open-sourced Qwen-Live-Delivery-Plugin for revising long-running agent work.
Timeline
dated events in the reporting · oldest first- Nov 16, 2026Alibaba releases the Qwen4-Max flagship model · CSO Online
What each outlet adds
beyond the shared facts- Hacker NewsWork360 scores, native vision, architecture, and no pre-training loss spikes.
- CSO OnlineSafeguard benchmark, token cut, planned open weights, and prompt retention.
- The DecoderIntelligence Index, exam score, cache pricing, and live-delivery plugin.
- Vendors
- Alibaba
Coverage
oldest first · one row per article · UTCQwen team releases Qwen4-Max, a 2.4-trillion-parameter model
It adds native vision, a one-million-token window, and 71B active parameters, with lower API prices than Qwen3.8-Max.
Alibaba releases Qwen4-Max with 2.4T MoE and 1M context
The title also lists 71B active parameters; the retrieved page had no further release details.
Alibaba releases Qwen4-Max and claims 62% lower task costs
The 2.4-trillion-parameter model activates about 71 billion parameters per token and supports a 1-million-token context.
Alibaba releases Qwen4-Max, priced below Opus 5.5, GPT-6 and Kimi
The 2.4-trillion-parameter mixture-of-experts model activates 71 billion parameters per token and offers a one-million-token context window.