New12,480+ smartphones indexed — free sandbox tier on every plan.Edge-cached JSON responses, typically under 100ms.SDKs for C#, TypeScript and Python, plus an MCP server for AI agents.New12,480+ smartphones indexed — free sandbox tier on every plan.Edge-cached JSON responses, typically under 100ms.SDKs for C#, TypeScript and Python, plus an MCP server for AI agents.
Get Free API Key
Artificial Intelligence & Frontier Models•4 min read•2026-11-16

Alibaba Releases Qwen4-Max Flagship Model: 2.4T Sparse MoE, 71B Active, and 1M Context

Alibaba unveils Qwen4-Max in general availability on November 16, 2026, pairing a 2.4T sparse MoE backbone (71B active) and hybrid Gated DeltaNet attention with $1.50/$4.50 per 1M token pricing.

Tech News Bot
DeviceSpecs Intelligence
Original Source Reference
#Alibaba#Qwen4-Max#Sparse MoE#AI Agents#Multimodal AI#1M Context#Artificial Analysis#QwenCloud

Alibaba has officially released Qwen4-Max into general availability, introducing its next-generation flagship multimodal AI model built on a 2.4-trillion-parameter sparse Mixture-of-Experts backbone. Activating 71 billion parameters per token alongside a 1-million-token context window across text, image, and video inputs, Qwen4-Max targets long-horizon autonomous agent workflows at substantially lower API rates than Western frontier rivals.

2.4T Sparse MoE Architecture with Hybrid Gated DeltaNet Attention

Qwen4-Max marks a major architectural shift from Qwen3.8-Max, combining a 2.4-trillion-parameter MoE backbone (71B active per token) with an additional 328 billion N-gram embedding parameters for low-compute local token lookups. To make 1-million-token context windows economical, each Qwen Sparse Attention (QSA) layer is paired with three Gated DeltaNet (GDN) layers that compress earlier history into fixed-size memory, stabilized by four parallel Gated Residual branches.

Benchmark Performance, Live Delivery, and Long-Horizon Agents

Designed around multi-hour autonomous execution, Qwen4-Max introduces Live Delivery, allowing long-running agents to ship usable initial outputs early and iterate while users redirect them mid-run. Independently, the model scored 43 on the Artificial Analysis Intelligence Index at $3.27 per evaluation task, while leading Kimi K3.1 on APEX-Agents v1.1 (72.1 vs 65.5) and JobBench (68.5 vs 61.2), though trailing top Claude and GPT-6 models on aggregate intelligence rankings.

Aggressive $1.50/$4.50 API Pricing and Closed-Weight Flagship Tier

Available immediately via QwenCloud and third-party routers like OpenRouter under the qwen4-max model ID, Alibaba priced Singapore/International API access at $1.50 per million input tokens and $4.50 per million output tokens ($1.413/$4.24 in Global regions), with cache reads at $0.075 per million. While Alibaba confirmed that Qwen4-Max will remain closed-weight, the company plans to release smaller open-weight models in the Qwen4 family at a later date.