Realtime AI News
Alibaba's Qwen4-Max debuts with 2.4T total parameters and 71B active, halving token use
Alibaba's Qwen team has released Qwen4-Max, a model with 2.4 trillion total parameters and 71B active per inference pass, with reports saying it consumes about half the tokens of its predecessor. Coverage also claims a 62% lower cost per task than Qwen3.8-Max while outperforming it.

Alibaba's Qwen team has released Qwen4-Max, and the headline number is its parameter split: 2.4 trillion total parameters against 71B active per inference pass. Chinese tech outlet 驱动之家 placed that combination in the same headline as a claim that token consumption falls by roughly half, which points at one idea — a bigger model that is cheaper to run.
The comparison with the previous generation is the second thread in the coverage. Startup Fortune reported that Qwen4-Max uses about half the tokens of Qwen3.8-Max on comparable tasks, cutting cost per task by 62%, and said the new model outperforms its predecessor.
The gap between 2.4 trillion total parameters and 71B active parameters indicates a sparse-activation design: the model holds a very large weight pool but routes only a small slice of it through each forward pass. That keeps compute and memory growth from scaling linearly with total size while still letting capacity grow.
For builders, the 62% figure matters more than the parameter count. A lower cost per task changes what agent products can afford: a long chain of calls and retries that was uneconomic at the old price becomes viable at roughly a third of it.
The cadence is also a signal. Qwen3.8-Max to Qwen4-Max is a short interval, and the improvement is aimed at efficiency rather than raw scale, matching what leading labs have done over the past year as capability gains flatten and cost per useful token becomes the competitive edge.
One caveat: the candidate material here is limited to media headlines and summaries. There is no official model card, pricing page, or statement of availability in the sources, so context length, rate limits, API pricing, and whether downloadable weights ship still need to be confirmed against Alibaba's own documentation.
Three things to watch: whether third-party benchmarks reproduce the claim of beating Qwen3.8-Max; whether API pricing lines up with the 62% cost-per-task figure; and how token efficiency holds up in agent, coding, and long-document workloads.
Why it matters
If the 62% cost-per-task reduction holds on real production traffic, the unit economics of agent and long-chain automation products get recalculated, and competition shifts further from raw parameter counts toward cost per completed task. Official pricing and independent benchmarks are needed to confirm the figure.
Nearby Updates
All11/16, 09:30
Apple launches a study app with OpenEvidence to test AI answers on health questions
Apple has launched a standalone app for a research study on how people use AI for health, built with healthcare AI company OpenEvidence, so participants can ask everyday medical questions and get answers grounded in clinical literature. The study is open to U.S. users over 21 on iPhone 15 or later with Apple Intelligence enabled, and Apple aims to enroll up to 200,000 people for more than a year.
11/16, 09:27
Anthropic's record IPO faces big questions before it reaches the market
France 24 reports that Anthropic's record-breaking IPO is heading toward the public markets while still facing a set of big questions. Rather than treating the listing as a settled triumph, the report frames the offering as a test of what a leading AI company must prove once its valuation is priced daily by public investors.
11/16, 08:00
AI Day Seoul wraps with 2,000-plus attendees focused on physical AI and agents
Nvidia's event recap says AI Day Seoul drew more than 2,000 attendees to COEX over two days, bringing together Korean industry leaders, startups and developer communities around physical AI and AI agents. At the plenary, Nvidia robotics and edge AI vice president Deepu Talla said Korea is punching above its weight in the field.
11/16, 07:30
Indian chip startup Bharat Semi raises $24M seed to target global AI sensor market
Bharat Semi, an Indian chip startup, has raised a $24 million seed round as it targets the global market for AI-enabled sensors. The company says it has secured design wins for its AI sensor technology with customers including General Atomics.