Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

Qwen4-Max is here: 2.4T parameters, a 1M-token context, and prices 25% below its predecessor

Alibaba released Qwen4-Max on 16 November 2026: a 2.4T-parameter flagship with a 1M-token context, API prices 25% below Qwen3.8-Max, and much stronger long-task ability. The catch: unlike its predecessor, Max ships with closed weights.

A developer working late at night at a desk with a laptop and an external monitor, both showing code

In short: Alibaba has released the first flagship of Qwen4 — cheaper, but the weights stay closed

On 16 November 2026, Alibaba's Qwen team announced Qwen4-Max on its official blog, the first model of the Qwen4 generation, pitched as "your new AI teammate for work, creativity, and discovery". It is a mixture-of-experts model with a 2.4-trillion-parameter backbone, a one-million-token context window, and API prices a quarter lower than its predecessor, Qwen3.8-Max, with cached input 70% cheaper. But one change deserves to come first: the previous generation's Max model shipped with open weights, and this one does not — for now you can only use it through Alibaba's cloud, and the weights are not available to download. Qwen only promises that "smaller models in the family" will be open-sourced; it says nothing about whether Max will follow.

First, keep two dates apart, because they are easy to confuse. What went live on 11 November was Qwen4-Max-Preview, a preview available only through the Token Plan subscription, which Alibaba's documentation says will be retired after 24:00 on 25 November (Beijing time). What is new on 16 November is the full release: model name qwen4-max, with a public pay-as-you-go API live the same day. The prices, benchmarks and features in this article are all for the full release. If you are already using the preview, switch before the 25th — don't wait until retirement day to discover your calls are failing.

One honest note on dating: the Qwen blog post itself carries no visible publication date stamp. The 16 November date is confirmed by cross-checking three independent sources — Model Studio's documentation was updated on 16 November with Qwen4-Max pricing and parameters, several independent writers who checked the blog and the pricing pages line by line (including System Design Academy) dated the release to 16 November, and Reuters reported on it the same day. That chain is solid, but the fact that the blog page itself shows no date stamp is recorded in the evidence file.

Start with the bill: 25% cheaper in and out, 70% cheaper cached input

For most developers, the most concrete change in Qwen4-Max is not in the parameter table but on the bill. Pay-as-you-go prices in Singapore are $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 per million implicitly cached input tokens. Qwen3.8-Max in the same region costs $2, $6 and $0.25. So input and output are each 25% cheaper, and cached input is 70% cheaper. On Alibaba's official price list, a request sending 100,000 tokens and receiving 10,000 costs about $0.195 before caching. Model Studio also offers a free quota of one million tokens, valid for 90 days from the latest of Model Studio activation, model release or application approval.

But the price cut is only half the story. Qwen says the new model also uses fewer tokens to do the same work: on Work360, its own evaluation of industry tasks, token use fell by about 24% at the highest reasoning effort and about 35% at high effort; on Terminal-Bench 4.0, the reductions are about 48% to 49%. Combined, the average cost per Work360 task fell from $4.43 to $1.67, a 62% reduction. For people running agents, the second number matters more than the first: an agent's bill is dominated not by expensive single calls but by the same long prompt and the same documents being resent at every step. Cached input at 70% off hits exactly that pain point.

One caution, though: Work360 is Qwen's own benchmark, and the per-task costs are Qwen's own calculations. A vendor's own benchmark numbers should always be taken with a discount — the only trustworthy verification is the cost of your own tasks. Independent evaluator Artificial Analysis scored Qwen4-Max 43 on its Intelligence Index on launch day (at the highest reasoning setting), 9 points above Qwen3.8-Max's 34, below Kimi K3.1 Pro's 46 and Claude Opus 5.5's 59; by its measure, one index task costs $3.27 against $10.02 for Qwen3.8-Max. The independent score and the vendor score point in the same direction, but the absolute gaps are very different — which is exactly why you should never trust a single source's numbers.

There is one more pricing uncertainty to state plainly. Some reports say these prices are introductory and could rise from 1 December, but that claim comes from second-hand reporting, and Alibaba's official price list says nothing about it. Until Alibaba confirms it, it is a rumour, not a fact to budget against. The safe approach is to budget short-term on the current price list and avoid locking long-term usage to a single vendor.

Three architecture changes: fewer active parameters, a bigger lookup table

Qwen4-Max has the same total size as its predecessor — 2.4 trillion parameters — but only 71 billion are active per token, down from 95 billion, about 25% fewer. At the same time it adds a 328-billion-parameter N-gram Embedding lookup table. In other words, the model did not get bigger; it spends its compute more frugally, with fewer parameters doing the work and more capacity placed in "lookup" operations, which are cheap. That is the technical basis for the price cut: inference got cheaper to run, so there was room to lower the price.

The official blog describes three specific architecture changes. First, attention: every Qwen Sparse Attention layer is paired with three Gated DeltaNet layers, which compress history into fixed-size states, while a lightweight indexer picks out the relevant parts of a long text at micro-block granularity. The point is to make a one-million-token input affordable in compute. Second, Gated Residual: four parallel residual branches use dynamic read and write gates to control information flow between layers; Qwen says this made training at the 2.4-trillion-parameter scale more stable, with no loss spikes throughout the run. Third is the N-gram Embedding mentioned above: a context-based lookup table at the second layer that stores patterns of nearby tokens, adding capacity for little extra computation.

In-house results for the base model: across Knowledge, Code, QA and STEM, Qwen4-Max-Base achieves 5% to 13% lower bits per byte (BPB — lower is better) than Qwen3.8-Max-Base, while activating 25% fewer parameters. BPB is a low-level metric: it says the model got better at predicting text, not that it will be better on your task. But it at least shows the improvement did not come from piling on parameters — the architecture itself is more efficient.

One caveat: Qwen has not published a technical report for Qwen4-Max, and the architecture section of the blog is only three short paragraphs. That is the full extent of the public information so far, with no paper to cross-check against. Anyone wanting deeper architecture detail will have to wait for a technical report or the open small models.

How to read the benchmarks: better than its predecessor across the board, mixed against everyone else

Qwen's blog includes a 16-test comparison chart putting Qwen4-Max, Qwen3.8-Max and four other models side by side. Against its predecessor the gains are clear: DeepSWE (agentic coding) 75.7 against 69.3, Terminal-Bench 4.0 (coding in a terminal) 48.5 against 38.9, Agents' Last Exam (long, high-value tasks) 46.7 against 27.9, and Work360 (industry tasks) 52.69 against 47.74. Across the eight tests we sampled, Qwen4-Max beats its predecessor on every one, with the biggest jump on Agents' Last Exam, nearly 19 points. The clearest progress is on long tasks — carrying a complex job through from start to finish, not just answering single questions well.

Sideways, the picture is less one-sided. In the same chart, Qwen4-Max leads all rivals on DeepSWE and Agents' Last Exam, but on Terminal-Bench 4.0 it scores 48.5 against Kimi K3.1's 67.2, a gap of nearly 19 points; on Humanity's Last Exam (hard questions across many subjects) 51.7 against Claude Opus 5.5's 59.8; and on Chartography (reading charts) 54.2 against GPT-6 Astra's 71.1. In other words, it is first-tier for long agent runs and industry work, but still behind on terminal coding and hard reasoning. Check your own task type before choosing: for long workflows and industry jobs, Qwen4-Max is worth trying; for pure terminal coding ability, it is not yet the strongest.

Qwen4-Max vs Qwen3.8-Max on four benchmarks: DeepSWE 75.7 vs 69.3, Terminal-Bench 4.0 48.5 vs 38.9, Agents' Last Exam 46.7 vs 27.9, Work360 52.69 vs 47.74; API input price $1.50 vs $2, output $4.50 vs $6 per million tokens

Read this chart with one eye open: it is Qwen's own chart, with Qwen's own measurements. The blog's footnotes say some scores are cited from public leaderboards — Agents' Last Exam from that test's own leaderboard, AutomationBench-AA from Artificial Analysis — but most are vendor-run. A vendor benchmark is not fabrication, but test conditions, prompts and reasoning-effort settings all move the results. Artificial Analysis's same-day score of 43 is currently the only independent anchor: it confirms the direction — better than its predecessor — but not the size of every lead in the vendor chart.

The four long jobs the blog spends most of its length on: games, law, medicine and making a new material

The most interesting part of this blog post is not the parameter table but the four examples of "long work" it spends most of its length on. Qwen clearly wants to send a signal: Qwen4-Max's selling point is not chatting, but independently finishing a complex job. The four examples have different levels of credibility, so they are worth taking one at a time.

The first is games and 3D modelling: the model built a cat-mage adventure game, "Meow of Magic", in Unreal Engine, a mobile racer set around West Lake in Hangzhou in Unity 6, and 3D architecture, interior and science-fiction scenes in Blender. The blog says the model runs the game itself during development and refines camera work, animation and pacing from visual and interaction feedback. These are single demonstrations, not "reproducible every time" capabilities. Whether the games are playable, how good the code is, and how many attempts it took — the blog says none of that. Treat it as a demonstration of the capability ceiling, not as a ready-to-use game development tool.

The second is legal due diligence, and it comes with the most specific numbers: 198 inconsistently formatted and named working papers, which the model filtered for noise, sorted into 14 folders by substantive content, indexed automatically, then cross-checked against public records covering business registration, intellectual property, judicial records and credit records, finally delivering an 86-page report of more than 60,000 Chinese characters, covering 12 review areas and 14 specific legal issues, with every conclusion backed by a working-paper citation. Specific numbers make it relatively more credible — but note this is a case Qwen's own team selected and presented, with no disclosure of document quality or how much human review was involved. Read it as a demonstration of the upper limit of long-document handling.

The third is critical care medicine: a complex case in which a patient developed ventricular fibrillation after carotid artery stenting, was resuscitated, then developed chest pain after extubation and needed reintubation on day 6. In a single conversation, the model read dozens of lab reports, more than 10,000 Chinese characters of progress notes, multiple sets of physician orders and 1,433 original medical images, organised six days of complex events into a coherent timeline, and produced four management recommendations that matched the attending physician's actual orders that day, point by point. This is the most striking example — and the one that most needs a cool head. Medicine is a high-stakes setting, and the blog's own ethics statement says the model assists work rather than replacing professional judgement. "Matching the doctor point by point" is a single case, not clinical validation.

The fourth made the biggest splash: a continuous 48-hour agent run in which the model autonomously designed and synthesised a novel coolant molecule on Deep Principle's Mira platform. Starting from a candidate space of 2.7 million molecules, it went through structural screening, property prediction, first-principles calculations, retrosynthetic analysis and experiment planning, then directed an automated wet lab to carry out synthesis, separation and purification, and testing — ending with a real physical sample of the material. The drug-discovery example is similar: targeting the NLRP3 inflammasome, 454 tool calls, 3,239 computational metric evaluations and 24 candidate molecules optimised, with the final molecule's nominal binding affinity estimated at about 3.6 times the reference molecule, endorsed by the partner's experts. The coolant produced a physical sample — that is concrete. The drug molecule is still at the computational stage; the partner says it "warrants further wet-lab validation", which means it has not been validated yet. The two should not be conflated.

One more industrial number deserves its own mention: in automotive manufacturing, the model helped plan real production lines and design parts, cutting the manual effort to program a single workpiece from about 475 minutes to 24 minutes, a reduction of about 95%. That figure comes from a real Alibaba Cloud business scenario, but again it is a single showcase, not a statistical average. Taken together, the four examples point to one judgement: Qwen4-Max's long-task ability is real, but the evidence so far consists of vendor-selected individual cases, not independent replication.

Safety and open source: two tightenings, one door left open

Safety is the second-biggest theme in the blog after performance, and it was also Reuters' angle on launch day. Qwen says it added code security checks on top of functionality checks in software-engineering training: on Last Secure Code Bench, the function-level pass rate for both functionality and security rose from 22.0% for the previous generation to 31.8%; the share of code generated securely at all three granularities — function, file and repository — rose from 8% to 14.7%, nearly doubling. The model does particularly well at avoiding core risks such as injection, memory safety and broken access control. The absolute numbers are still low — 31.8% means nearly seven in ten cases still fail to satisfy functionality and security together — but the direction is right, and Qwen deserves credit for publishing a "how much of the code we generate by default is vulnerable" metric at all.

On cybersecurity and CBRN (chemical, biological, radiological, nuclear), the blog's wording is much more cautious: training data is filtered by the purpose of each task, keeping data for defensive work such as vulnerability verification, detection and remediation, and removing data aimed at enhancing attack capabilities; an internal expert review of long-horizon task trajectories found substantially improved defensive capability and strengthened safety against offensive uses, with no capabilities observed that pose large-scale cybersecurity risks; on CBRN, the model identified more potentially risky requests and gave safe responses, and expert review found the solutions it proposes are "not genuinely innovative and cannot complete the task end-to-end". These are internal evaluations with no independent third-party backing, and the wording leaves room. Read them as the vendor's own safety account, not as an independent audit conclusion.

On open source, Qwen's position is clear: Max itself is not open, and its weights are only available through the cloud; "smaller models in the family are coming to open source", built on the same lineage and trained with the same generation of techniques, but with no timetable. That contrasts with the previous generation: the weights of Qwen3.8-Max, Qwen3.8-2.4T-A95B, went up on Hugging Face in August 2026, the first time Qwen opened the weights of a Max-class model. This generation has pulled back. The blog does not say why, but commercially it is easy to understand: the flagship model is the funnel into the cloud service, and open small models with a paid flagship is now the standard playbook. For developers, the practical consequence is that local deployment, fine-tuning and offline use are off the table for Qwen4-Max; wait for the small open models, or keep using the previous generation's open-weight version.

One open-source move is easy to miss: Qwen open-sourced the Live Delivery plugin. Live Delivery is the long-task mechanism described in the blog — a task does not wait until it is fully finished before delivering, but hands over usable versions in stages, with the user able to step in and correct direction at any point. Open-sourcing the plugin means the mechanism is not limited to Qwen's own products; other agent builders can adopt it. For developers building agent products, this plugin may be more worth a look than the model itself.

How to try it, and whether to switch: run your own numbers before moving production

Getting started is straightforward. The API runs on QwenCloud under the model name qwen4-max, and it supports the OpenAI and Anthropic API formats, so existing code can usually be pointed at it by changing the base address and the model name; it is also available through the Token Plan subscription, or directly in QwenWork, Qoder and QwenStudio. The context window is one million tokens; QwenCloud lists a maximum input of about 991,000 tokens (about 983,000 with thinking on) and a maximum output of about 131,000 tokens (about 262,000 for thinking plus the final answer combined). Reasoning is on by default, with effort adjustable from low to xhigh; inputs accept text, images and video, and the model supports function calling, structured JSON output, web search and batch requests.

Whether to switch depends on the shape of your bill. If your agents spend every day re-sending the same long prompt and the same large documents, 70%-cheaper cached input plus lower token use will add up to visible savings; if your calls are mostly short prompts, the 25% price cut is simply a 25% cut, with no extra surprise. If your work is agentic coding in a terminal, the benchmarks show it still trails Kimi K3.1, so run a comparison on your own repositories before switching. If you care about local deployment and fine-tuning, it is not in scope at all right now — wait for the small open models.

Read in context: not a bid for "strongest", a bid for "cheapest long-task partner"

Placed in November's model landscape, Qwen4-Max's positioning is clear: it is not competing for the "strongest model" title, but for the "most cost-effective long-task partner" slot. In the same comparison chart, Claude Opus 5.5 scores 59.8 on Humanity's Last Exam, Kimi K3.1 scores 67.2 on Terminal-Bench, and GPT-6 Astra scores 71.1 on chart reading — someone beats it on every single test. But Qwen4-Max plays a different game: a million-token context, long tasks that don't fall apart, prices a quarter lower than its predecessor, and cached input 70% cheaper. For a team running dozens of agent tasks a day, "good enough on everything, cheapest on the bill" is more attractive than "first on one test".

That positioning also explains why the blog spends so much space on law, medicine and materials: Qwen wants the market to remember not a benchmark score but "it can carry a big job through from start to finish". That is a different pitch from Claude's emphasis on coding or Kimi's on terminal coding. Whether the differentiation holds will depend on independent replication: so far every long-task case is one Qwen selected itself, and Artificial Analysis's score of 43 covers standard benchmarks, not a 48-hour coolant run. The positioning is only secure once third parties publish comparisons on real tasks.

All told, Qwen4-Max is a pragmatic release: no parameter race, no "strongest model" slogans, just three messages — cheaper, better at long work, and not open for now. For Alibaba, it is the opening of the Qwen4 generation and a funnel into its cloud service; for developers, its value is not in launch-day surprise but in your own bills and task records. Benchmarks and demos can only tell you "it might work"; only your own side-by-side runs can tell you whether it works for you. One last practical note: preview users should switch from qwen4-max-preview to qwen4-max before 25 November, and not discover the retirement the hard way.

Views 4Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project