cd /news/large-language-models/alibaba-s-qwen4-max-beats-qwen3-8-ma… · home › topics › large-language-models › article
[ARTICLE · art-170225] src=startupfortune.com ↗ pub= topic=large-language-models verified=true sentiment=↑ positive

Alibaba's Qwen4-Max beats Qwen3.8-Max with half the tokens and a 62% lower cost per task

Alibaba's Qwen team announced Qwen4-Max, a mixture-of-experts model with native vision and a 1M-token context window that Alibaba says outperforms Qwen3.8-Max using up to roughly 50% fewer tokens, with average cost per task falling more than 60%. According to AlphaSignal and RuntimeWire, Alibaba puts average task cost on its Work360 benchmark at $1.67 for Qwen4-Max versus $4.43 for Qwen3.8-Max, a cut of about 62%, helped by QwenCloud pricing of $1.50 per million input tokens and $4.50 per million output tokens, each down 25% from $2 and $6, and implicit caching at $0.075 per million tokens, a 70% cut. Alibaba says Qwen4-Max scores 52.69 on Work360, which it describes as covering 721 professional fields and which AlphaSignal calculates as roughly 10% above Qwen3.8-Max, though the first-party benchmark has not been independently reproduced and Alibaba has not disclosed parameter counts or the expert setup.

by read4 min views4 publishedNov 16, 2026
Alibaba's Qwen4-Max beats Qwen3.8-Max with half the tokens and a 62% lower cost per task
Image: Startupfortune (auto-discovered)

Alibaba says its new flagship does more with far less, and that the savings show up on the invoice, not only on a chart. Before you rewrite a budget around it, know that the headline numbers come from Alibaba's own test.

Alibaba's Qwen team announced Qwen4-Max on X as "your new AI teammate for work, creativity, and discovery," describing a model built on a new mixture-of-experts architecture with native vision and a 1M-token context window. The post says it outperforms Qwen3.8-Max with up to roughly 50% fewer tokens, while average cost per task falls by more than 60%.

That second figure has a number behind it. According to AlphaSignal and RuntimeWire, Alibaba puts the average cost of a task on its Work360 benchmark at $1.67 for Qwen4-Max, down from $4.43 for Qwen3.8-Max. That's a cut of about 62%.

Two things are stacking here, and it helps to separate them. The first is that the model needs fewer tokens to finish the same job. Alibaba says it uses as much as 50% fewer than Qwen3.8-Max across its evaluations, which matters because Qwen models have had a reputation for long, chatty outputs that inflate the bill. Artificial Analysis, which tracks Qwen3.8-Max, lists it at $5.41 per task on its own Intelligence Index, a figure that shows how far token habits can move cost even when the per-token price looks low.

The second is plain price cutting. AlphaSignal reports that Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens on QwenCloud, each down 25% from Qwen3.8-Max's $2 and $6. Implicit caching drops to $0.075 per million tokens, a 70% cut. Fewer tokens at a lower rate gets you most of the way to that 62%.

Qwen4 Nears Launch as Alibaba's Open Source Promise Faces Real Doubt Alibaba Cloud CTO Feifei Li told the Apsara Conference in Amsterdam on October 29 that Qwen4 could arrive within the week. Qwen models have passed 700 million Hugging Face downloads, so developers building on them now want to know whether the weights will be open. - Alibaba Qwen4 open source model release date - Qwen4 context window and model family specs

On quality, Alibaba says Qwen4-Max scores 52.69 on Work360, a benchmark it describes as covering 721 professional fields, which AlphaSignal calculates as roughly 10% above Qwen3.8-Max.

What you can and can't check yet #

Frankly, treat the Work360 numbers as a claim, not a result. Work360 is Alibaba's own benchmark, and AlphaSignal notes that its first-party origin limits how much weight the comparison can carry until independent evaluators reproduce it. Nobody outside Alibaba has published a rerun that we've found.

The architecture is also thinner on detail than the post suggests. AlphaSignal reports that Alibaba hasn't disclosed total or active parameter counts or the expert setup, so the "new MoE architecture" line can't be tested from the outside. Some third-party write-ups quote specific parameter figures, but we couldn't tie them to anything Alibaba has said, so we're leaving them out.

The rollout has been messier than a single launch day. Qwen4-Max went into preview on November 11 on Alibaba's Token Plan, with preview access running through November 25, according to Xenospectrum. AlphaSignal says it's now reachable through the QwenCloud API, Qwen Studio, QwenWork and Qoder. The same report says Alibaba has promised smaller open-weight Qwen4 models but hasn't given a date.

If you build on Qwen3.8-Max today, the test is easy to describe and cheap to run. Take your ten most expensive real workloads, run them on both models, and compare cost per finished task, not price per token. A model that talks half as much can beat a cheaper one that doesn't, and the reverse is just as possible on your data. Alibaba's $1.67 against $4.43 is where to start the comparison, not where it ends. The 1M-token window comes with a limit worth knowing about. Xenospectrum's reading of the preview is that the base context is 262K tokens, extensible to 1M, so very long inputs may not behave like the headline number on day one.

Also read: AI child abuse reports have jumped 8,400% in two years, Childlight data shows • Meta's 'Rapture' payout in February could let some veteran staff retire early • How to Version Prompts for AI Agents Before a Bad Edit Hits Production

Alibaba's CEO Says Its Next AI Model Could Reach 10 Trillion Parameters Alibaba CEO Eddie Wu announced plans to train an AI model with 5 to 10 trillion parameters, alongside a new Zhenwu V900 chip and a 20-gigawatt data center target by 2032. The announcement landed two days before Trump and Xi meet in Washington, where AI is a top agenda item. - alibaba's 10 trillion parameter AI model plans - how large will alibaba's next AI model be

This article is posted in AI News, check it out for more related stories.

Join the discussion #

Open in the community → Almost there. Sign in and your reply posts straight away.

── more in #large-language-models 4 stories · sorted by recency
── more on @alibaba 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/alibaba-s-qwen4-max-…] indexed:0 read:4min 2026-11-16 · —