Go to PegaPoll
PEGAPOLL · NEWS

Alibaba’s Qwen4-Max targets frontier rivals at sharply lower prices

The Decoder (AI) · 2026-11-16
🤖 AI-generated content — The title and summary were produced automatically by artificial intelligence, without human editorial review.

Alibaba has launched Qwen4-Max, a mixture-of-experts flagship with 2.4 trillion parameters that activates only 71 billion per token. It processes images natively, offers a one-million-token context window, and is aimed at professional work in industry, science and finance, at prices meant to undercut rivals such as Opus 5.5, GPT-6 and Kimi K3.1 Pro. Through Alibaba’s API, input tokens cost $1.50 per million, output $4.50, and cache hits $0.075.

The Qwen4 architecture mixes compressed context with selective attention: three layers condense prior text and a fourth attends only to what matters, cutting compute on long inputs. Four parallel connections with dynamic gates sit between layers, and an N-gram embedding of 328 billion parameters supplies common token sequences. Training of the 2.4-trillion-parameter setup was described as stable. Internally, Qwen4-Max-Base led five checkpoints, improving 5 to 13 percent over Qwen3.8-Max-Base while activating 25 percent fewer parameters per token.

On Alibaba’s Work360 industrial benchmark covering 721 fields, the model uses about a quarter to a third fewer tokens than its predecessor; on Terminal-Bench 4.0, token use nearly halves. Average Work360 cost fell 62 percent, from $4.43 to $1.67, while the score rose from 47.74 to 52.69. External results are mixed: 46.7 on Agents’ Last Exam v1.1 (best in the group), behind Opus 5.5 and GPT models on Humanity’s Last Exam, and past Opus 5.5 but behind GPT on HealthBench Professional. At its highest reasoning setting it scores 43 on the Artificial Analysis Intelligence Index—nine above Qwen3.8-Max and three behind Kimi K3.1 Pro at 46—at $3.27 per task versus $10.02 and $3.68. Tencent’s Hy4 is cheaper at $1.03 but eight points lower. On Terminal-Bench 4.0 Qwen4-Max scores 48 percent against Kimi’s 67 percent; on FrontierCode 2.0 it ties Kimi at 26 percent and trails Hy4 at 31 percent. Alibaba open-sourced a Qwen-Live-Delivery-Plugin for staged agent output and plans smaller open-source Qwen4 models. Partners SANF and Pacific Union reportedly cut workpiece programming from about 475 minutes to 24.

Continue in the app — vote & join in ➔
Source: The Decoder (AI) · via ahirlevel.hu

Related news