Ai Engineering 6 min read

Qwen4-Max: Alibaba's 2.4T closed model cuts prices and claims a 62% cheaper task, but trails US rivals on independent tests

Alibaba released Qwen4-Max on November 16 with a new Gated DeltaNet plus sparse-attention architecture, 71B active parameters, 1M context and $1.50/$4.50 pricing. Its own Work360 benchmark shows a 62% lower cost per task; Artificial Analysis puts it at 43, behind Kimi K3.1 Pro.

Alibaba’s Qwen team released Qwen4-Max on November 16, the first model of the Qwen4 generation and the company’s new flagship. The launch post pitches it as an “AI teammate” for long-running professional work, built on a new architecture with a 2.4-trillion-parameter backbone, native vision and a one-million-token context window. The model is closed-weight; Alibaba says smaller Qwen4 models “are coming to open source” without giving a date.

The headline claim is economic rather than raw capability: on Alibaba’s own Work360 benchmark, the average cost per task fell from $4.43 with Qwen3.8-Max to $1.67, a 62% reduction, while the score rose. Independent leaderboards published the same day are more modest about where it sits against US frontier models and against China’s strongest rival, Moonshot’s Kimi K3.1 Pro.

A new architecture, not a bigger Qwen3.8

Qwen3.8-Max, which we covered in August, was already a 2.4T-class mixture-of-experts model. Qwen4-Max keeps the backbone size but changes most of what sits around the experts. Per the launch post, it activates 71B parameters per token, 25% fewer than Qwen3.8-Max-Base, plus an additional 328B “N-gram Embedding” parameters. The three main changes:

  • Gated DeltaNet plus Qwen Sparse Attention (QSA). For every QSA layer there are three Gated DeltaNet layers, a linear-attention design that compresses history into a fixed-size state. QSA then uses a lightweight indexer to pick relevant context at “micro-block granularity.” The goal is to make the 1M-token window affordable: most layers do not grow with sequence length, and the ones that do only read a selected subset.
  • Gated Residual. Four parallel residual branches with dynamic read and write gates, which Alibaba says strengthens information flow across layers and stabilizes training.
  • N-gram Embedding. Context-based lookups at the second layer that represent local token patterns, adding capacity “with little additional computation.” This is why the parameter count is quoted as backbone plus embedding rather than a single number, and why some coverage describes the model as 2.4T total while the blog describes 2.4T plus 328B.

Alibaba reports no loss spikes across the full pre-training run and says Qwen4-Max-Base achieves 5-13% lower bits per byte than Qwen3.8-Max-Base across knowledge, code, QA and STEM evaluations. Those are in-house measurements on base models, and Alibaba notes that its comparison checkpoints for GLM-5.3 and Kimi K3.1 Pro were post-trained models, which can skew the comparison.

Where the 62% comes from

The cost figure has three components. Token use fell: on Work360, Qwen4-Max used about 24% fewer tokens at Xhigh effort and 35% fewer at High, and on Terminal-Bench 4.0 the reductions were about 48% and 49%. Prices fell: international list pricing reported by DataCamp and MLQ is $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6, with cached input at $0.075 per million. And Alibaba says cached-input pricing saw the largest cut, which matters for agentic workloads that resend long contexts on every step.

DataCamp points out that the token and list-price cuts alone imply roughly a 51% saving, not 62%; the remainder presumably comes from caching behavior on long tasks, but Alibaba has not broken it down. Work360 itself is Alibaba’s benchmark, built on a taxonomy of 26 domains, 182 subdomains and 721 specialized fields. Qwen4-Max scored 52.69 against Qwen3.8-Max’s 47.74. No outside group has reproduced it.

What independent evaluations say

Third-party numbers tell a more mixed story. Trending Topics reports a score of 43 on the Artificial Analysis Intelligence Index, a 9-point gain over Qwen3.8-Max but behind Claude Fable 5.2 (65), Gemini 4 Carbon (60), Claude Opus 5.5 (59), GPT-6 Astra (57) and Kimi K3.1 Pro (46). Artificial Analysis also measured it as verbose: 127 million output tokens to run the index against a median of 78.6 million, for an index cost of $3.27, about a third of Opus 5.5’s $10.04 but well above GPT-6.1 Sol’s $1.27.

On the Vals Index, Wccftech and DataCamp report 11th of 38 at 63.2%, the highest placement for a Chinese model in that table. Alibaba’s own comparison table, as compiled by DataCamp, shows the pattern clearly: Qwen4-Max leads Opus 5.5 on some agentic work benchmarks (21.3 vs 9.2 on HLAB-Vals-1.1, 46.7 vs 43.4 on Agents’ Last Exam v1.1) but trails on coding, with 48.5 on Terminal-Bench 4.0 against 59.6 for Opus 5.5 and 67.2 for Kimi K3.1 Pro. Vals’ own Terminal-Bench run measured 48.48%, which matches Alibaba’s figure.

The fair reading: Qwen4-Max is a large step over its predecessor and a strong value option for long professional tasks, but on independent general-intelligence and coding measures it does not overtake the US frontier or Kimi.

Cyber guardrails by data filtering

The safety section is notable because it describes a specific technique rather than a policy. For cybersecurity, Alibaba says it filtered training data “according to the specific purpose of each task,” keeping data for defensive work such as vulnerability verification, detection and remediation, and removing data intended to enhance attack capabilities. An internal expert review of long-horizon trajectories, including full chain of thought, found “no capabilities posing large-scale cybersecurity risks.” On CBRN, it added dedicated safety data in post-training and says reviewers judged the model’s proposed solutions “not genuinely innovative.”

Alibaba also added security checks to software-engineering training. On Last Secure Code Bench, the share of function-level outputs passing both functionality and security checks rose from 22.0% to 31.8%, and code that was secure at the function, file and repository level simultaneously rose from 8% to 14.7%. Those are improvements, but they also mean most generated code still fails the combined check. Digital Trends notes that no external red team has tested the model yet, so the cyber claims rest on Alibaba’s internal review.

Availability and what to watch

Qwen4-Max is available through the QwenCloud API (model ID qwen4-max) and Token Plan, and in Qwen Studio, QwenWork and the Qoder coding product. Alibaba also open-sourced a Qwen-Live-Delivery-Plugin, which implements the post’s “Live Delivery” pattern: long-running agents ship a usable first version early and iterate in releases the user can steer, rather than returning one result at the end.

Things to watch next:

  • The open Qwen4 models. Alibaba says they are built on the same lineage and techniques as Max. If the smaller Qwen4 models carry over the Gated DeltaNet plus QSA design, they could make 1M-token contexts much cheaper to self-host.
  • Independent cost-per-task checks. The 62% figure is the main selling point; reproductions on third-party agentic suites will show whether the token savings survive outside Work360.
  • External cyber evaluation. Training-data filtering is a cheaper alternative to post-hoc refusal training. A third-party offensive-capability evaluation would show whether it actually removes skills or only hides them.
Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Engineering

gr.Workflow Turns AI Pipelines Into Deployable APIs

Learn how to model, debug, expose, and deploy multi-step AI pipelines with Gradio Workflow and daggr.

Ai Engineering

Qwen Image 2.1 Ships Native Transparency and 10-Reference Editing in a 7B Open-Weights Model

Alibaba's Qwen team released Qwen Image 2.1, a 7B open-weight model that unifies image generation and editing, outputs native RGBA transparency at 2K, and runs on consumer GPUs. The catch is a research-only license.

Ai Engineering

Meta's Muse Ember matches Haiku 5.5 and GPT-6 Luna on list price and competes on tokens per task

Meta released Muse Ember on November 9 at $0.10 per million input tokens and $0.50 output, the same rates as Claude Haiku 5.5 and GPT-6 Luna. Its pitch is cost per task: about $0.14 on the Artificial Analysis index against $0.41 for Haiku at a similar score.

Ai Engineering

Claude Fable 5.2: 20% cheaper, a 2.5x fast mode, and fewer requests handed to older models

Anthropic's Fable 5.2 and Mythos 5.2 cost $8 and $40 per million tokens, beat Opus 5.5 on most published tests, and block about 55% fewer requests. What changes for developers, with the benchmark table and prices.

Ai Engineering

Editing 1,023 weights raised Qwen3.5-0.8B's MATH-500 score by 29 points. Most of the paper is about why that rarely works

A new paper from Shanghai AI Lab, Tongji and Fudan tests agent-written weight edits across 16 trajectories. The headline gain is real on held-out data, but six runs kept the original weights and the larger model barely moved.