DEV Community

Toàn Thành thy
Toàn Thành thy

Posted on

Qwen4-Max Is Here: A 2.4T-Parameter Teammate That Cuts Your Agent Bill by 62% — With a Catch

Alibaba's Qwen team dropped a new flagship today: Qwen4-Max, built on a brand-new Qwen4 architecture with a 2.4-trillion-parameter backbone, native vision, and a one-million-token context window. The headline number, though, is not the parameter count. It is the price: on Qwen's own Work360 evaluation, the average cost per task fell from $4.43 to $1.67 — a 62% cut against the previous generation, Qwen3.8-Max.

That is a big claim, and it is worth unpacking carefully, because the interesting part of this release is not any single benchmark. It is how the model was built to be cheaper to run and harder to derail on long, messy, real-world jobs — and where the announcement still leaves questions unanswered.

The architecture: big backbone, small active footprint

Qwen4-Max is a sparse mixture-of-experts model: 2.4T backbone parameters, but only 71B activated per token, plus an additional 328B of N-gram Embedding parameters. Three architectural changes do the heavy lifting:

  • GDN + Qwen Sparse Attention (QSA). Three Gated DeltaNet layers compress history into fixed-size states for every sparse-attention layer, and a lightweight indexer selects relevant context at micro-block granularity. The point is cheaper long-context attention — which matters when your context window is a million tokens.
  • Gated Residual (GR). Four parallel residual branches with dynamic read and write gates, aimed at stronger information flow and more stable training.
  • N-gram Embedding. Context-based lookups at the second layer give the model representations of local token patterns, expanding capacity with little extra compute.

The training run itself is part of the story: Qwen reports a smooth loss curve with no loss spikes throughout the 2.4T-parameter run, which it credits to the architecture and training optimizations. On in-house evaluations across knowledge, code, QA, and STEM, the Qwen4-Max base model achieved the lowest bits-per-byte of five evaluated checkpoints — roughly 5–13% lower than Qwen3.8-Max-Base while activating 25% fewer parameters.

Cheaper, not just stronger

The cost numbers are the most concrete part of the announcement. Against Qwen3.8-Max, Qwen4-Max uses fewer tokens on the same work: about 24% less on Work360 at Xhigh effort and 35% less at High effort, and about 48–49% less on Terminal-Bench 4.0. API prices have also been lowered, with particularly large cuts to cached-input pricing. Combined, that is what produces the $4.43-to-$1.67 drop in average cost per Work360 task.

Work360 itself is worth a note: it is Qwen's own framework for modeling complex professional work, with a taxonomy spanning 26 Tier-1 domains, 182 Tier-2 subdomains, and 721 Tier-3 fields. On its benchmark, Qwen4-Max scored 52.69, up from 47.74 for Qwen3.8-Max. A higher score on your own benchmark is expected; the more meaningful signal is that the gains came with lower token usage rather than more of it.

Built for long-horizon work, not just chat

The release is framed around a "24/7 AI teammate," and the demos lean hard into long-running, multi-step work:

  • 3D and games. The model drives Blender for geometry, materials, lighting, and rendering, and builds playable games in Unity, Unreal Engine, and Godot from a single brief — including running the game and iterating on camera tracking, animation, and pacing from visual feedback.
  • Manufacturing. In real production settings at SANF and Pacific Union, Qwen4-Max planned production lines, designed parts, and verified machining — cutting the manual effort to program a single workpiece from about 475 minutes to 24 minutes, roughly a 95% reduction.
  • Finance, law, and healthcare. For investment research it invokes 90 MCP tools across market data and produces Excel workbooks and PowerPoint decks where charts map one-to-one to the underlying data. For pre-IPO legal due diligence it sorted 198 inconsistently formatted working papers into 14 folders and produced an 86-page report of over 60,000 Chinese characters, with every conclusion backed by citations. For critical care it reviewed a six-day clinical course spanning dozens of lab reports and 1,433 medical images, and its four management recommendations matched the attending physician's actual orders point by point.
  • Live Delivery. Instead of making users wait for a long task to finish, the agent ships a usable first version early and improves it through successive releases, with the user able to step in and redirect at any point. Qwen is open-sourcing the Qwen-Live-Delivery-Plugin so others can build on the pattern.

Two science demos push the long-horizon claim furthest. In a continuous 48-hour run, Qwen4 surveyed approaches, screened 2.7 million candidate molecules, and directed an automated wet lab through four design–synthesis–testing loops to produce a physical sample of a novel coolant material for data-center chips. In drug discovery targeting the NLRP3 inflammasome, it made 454 tool calls and 3,239 computational metric evaluations across 24 candidate molecules, ending with a candidate whose estimated binding affinity is about 3.6 times the patent-derived reference molecule.

The catch: what the announcement does not prove

None of this is independently verified yet, and the announcement is candid about its own framing in ways worth noticing.

First, Qwen4-Max itself is not open source. The blog says smaller models in the family are "coming to open source" on the same lineage — Max is the hosted flagship, available through QwenCloud (Token Plan or API), QwenWork, Qoder, and Qwen Studio. If you wanted to self-host this one, that door is not open.

Second, the headline benchmarks are largely in-house. Work360 is Qwen's own benchmark, and the cost-per-task figures come from that same evaluation. Third-party scores cited in the footnotes come from leaderboards like LM Arena and Artificial Analysis, but the announcement leads with its own numbers. Treat the 62% as a directional claim about Qwen's own workload mix, not a universal price comparison.

Third, the most spectacular demos — a new coolant molecule, a drug candidate — are single showcases run with partner platforms (Deep Principle's Mira and DP Technology's PharmMaster). They demonstrate that a long-horizon agent loop can complete, not that it will complete reliably on your problem. The drug candidate, for instance, has expert endorsement but still awaits wet-lab validation.

Finally, the safety numbers are genuinely encouraging but modest in absolute terms: on Last Secure Code Bench, the function-level pass rate for both functionality and security rose from 22.0% to 31.8%, and code that is secure at all three granularities (function, file, repository) rose from 8% to 14.7%. Nearly doubling is real progress — and a reminder that most generated code still is not secure by default.

The bottom line

Qwen4-Max is a bet that the next frontier is not a smarter chatbot but a cheaper, steadier long-horizon worker: fewer tokens per task, a million-token memory, and an agent loop that delivers early and iterates in the open. The architecture — sparse attention with compressed history, gated residuals, and N-gram embeddings — is designed to make that economics work, and the training stability at 2.4T parameters is a real engineering result.

Whether the 62% cost cut holds on your workload is something only your own evaluation will tell. But if the model delivers even half of that efficiency on long agentic tasks, it changes the math for anyone running agents at scale — and the smaller open-source Qwen4 models coming next will be the ones to watch for self-hosters.


Sources: Qwen Team, "Qwen4-Max: Your New Teammate Has Arrived," Qwen Blog, November 16, 2026. All figures are as reported in the official announcement.

Top comments (0)