Qwen4-Max: 2.4T parameters, $1.50 input and no open weights
Alibaba's new flagship activates 71B of 2.4T parameters, costs $1.50/$4.50 per million tokens with cached input at $0.075, and is API-only. What the architecture, price and cyber claims mean for builders.
Alibaba released Qwen4-Max on 16 November 2026, and the first thing builders should know is what it is not: it is not an open-weight model. The previous flagship, Qwen3.8-Max, was the first Max-class Qwen whose weights Alibaba published. Qwen4-Max is available only as a hosted model, through Alibaba’s QwenCloud API and Token Plan subscription, its own apps and third-party gateways. Alibaba says “open-weight, smaller-scale models from the Qwen4 series will soon be released”, without naming them or giving dates.
What it offers instead is a lower price for a bigger claimed step. Alibaba describes a new mixture-of-experts architecture with a 2.4-trillion-parameter backbone and 71 billion parameters active per token, down from about 95 billion on Qwen3.8-Max, plus a 1-million-token context window and native vision. QwenCloud’s Singapore list price is $1.50 per million input tokens and $4.50 per million output tokens, 25 per cent below Qwen3.8-Max, and cached input falls 70 per cent to $0.075. Alibaba combines those prices with lower token use to claim a 62 per cent cut in average cost per task on Work360, a benchmark it built.
This article covers the architecture as Alibaba describes it, the full price list and how it compares, what the licence situation means for teams that self-host, the benchmark claims and the one independent aggregate measurement so far, and the cybersecurity safeguards. GenAI Brief has not run the model; vendor results are labelled as such.
What Alibaba released, and where
Alibaba published two primary documents on 16 November: a technical Qwen blog post, “Qwen4-Max: Your New Teammate Has Arrived”, and a shorter Alibaba Cloud press release datelined Hangzhou. The blog calls Qwen4-Max “your new AI teammate for work, creativity, and discovery” and positions it for long-running agent tasks across 3D modelling, game development, manufacturing, finance, law, healthcare and scientific research.
According to The New Stack, Alibaba teased the Qwen4 family the week before and offered Qwen4-Max in preview through several of its products. Vercel’s changelog added the model to its AI Gateway on 15 November under the ID alibaba/qwen4-max.
The access routes Alibaba lists are:
- QwenCloud API, pay as you go, with OpenAI- and Anthropic-compatible interfaces, under the model name
qwen4-max. - Token Plan, a QwenCloud subscription billed in credits for use inside coding and agent tools. The Token Plan documentation names Claude Code, Cursor, Qwen Code, Codex, Qoder and OpenClaw as compatible tools, and says subscriptions can be used only in the Singapore region. Its Team Edition “guarantees that your data will not be used for model training.”
- Alibaba’s own products: QwenWork, Qoder and QwenStudio.
The QwenCloud model page describes Qwen4-Max as “the next-generation flagship multimodal model in the Qwen family”. It takes text, image and video input and returns text. Its listed features are prefix completion, function calling, context caching, structured outputs, batch processing, web search and fine-tuning, with built-in tools including a code interpreter, web search and web extraction through the Responses API.
The architecture, as Alibaba describes it
The Qwen blog’s pre-training section gives the parameter counts: “2.4T backbone parameters, 71B activated per token, and an additional 328B N-gram Embedding parameters”. It names three architectural changes across “attention, residual connections, and embeddings”.
- GDN plus Qwen Sparse Attention. “Three Gated DeltaNet layers compress history into fixed-size states for every QSA layer.” The QSA layers use “a lightweight indexer to select relevant context at micro-block granularity,” which Alibaba says reduces attention and indexing costs on long sequences.
- Gated Residual. “Four parallel residual branches use dynamic read and write gates to strengthen information flow across layers and improve training stability.”
- N-gram Embedding. “Context-based lookups provide representations of local token patterns at the second layer, expanding model capacity with little additional computation.”
SiliconANGLE’s report describes the N-gram parameters as functioning as “a kind of cache” for oft-repeated calculations. That is the outlet’s gloss; Alibaba’s own description is the lookup of local token patterns quoted above.
Two claims about training are also Alibaba’s. The blog says pre-training at the 2.4-trillion scale ran with “no loss spikes throughout the run”, attributed to the architecture and optimisation changes. It also says Qwen4-Max-Base achieved about 5 to 13 per cent lower bits per byte than Qwen3.8-Max-Base across knowledge, code, question answering and STEM in-house evaluations “while activating 25% fewer parameters”. The comparison chart includes GLM-5.3 and Kimi-K3.1-Pro, and the caption notes that those “are post-trained checkpoints, which may affect comparisons with base models.”
Counting the parameters
The two Alibaba documents count differently. The press release says Qwen4-Max features “2.4 trillion total parameters” and activates 71 billion per token, “25% fewer activated parameters compared to Qwen3.8-Max at the same 2.4T scale”. The technical blog calls the 2.4 trillion the “backbone” and lists 328 billion N-gram Embedding parameters as “additional”. Read literally, the stored model is about 2.73 trillion parameters. The distinction would matter for anyone sizing hardware to host it, but since no weights are published, it matters mainly for comparisons with other labs’ “total parameter” figures.
The 25 per cent reduction in active parameters is consistent with the previous release. Our October survey of open-weight models recorded Qwen3.8-2.4T-A95B, published in August, as 2.4 trillion total and 95 billion active; 71 billion is about 25 per cent fewer. Fewer active parameters per token reduce the compute per generated token, which is the likely mechanism behind the lower list price, although Alibaba does not publish its serving costs.
Context and limits
QwenCloud’s model page lists the operating envelope in the Singapore region:
| Limit | Qwen4-Max | Qwen3.8-Max |
|---|---|---|
| Context | 1M | 1M |
| Max input | 991K | 991K |
| Max input, thinking mode | 983K | 983K |
| Max output | 131K | 131K |
| Max output, thinking mode | 262K | 131K |
| Max reasoning | 262K | 262K |
| Tokens per minute | 5,000,000 | 2,000,000 |
QwenCloud model pages, Singapore region, read 18 November 2026.
The practical changes are a doubled output ceiling in thinking mode and two and a half times the per-minute token allowance. Artificial Analysis lists a 984,000-token context window for the model.
The price list
QwenCloud’s pricing documentation gives Qwen4-Max its own table and states that “thinking and non-thinking modes use the same price, with no input-length tiers.” Thinking tokens are billed at the output rate.
| USD per 1M tokens | Qwen4-Max, Singapore | Qwen4-Max, Beijing / Global | Qwen3.8-Max, Singapore | Change, Singapore |
|---|---|---|---|---|
| Input | $1.50 | $1.413 | $2.00 | −25% |
| Output (thinking + answer) | $4.50 | $4.24 | $6.00 | −25% |
| Implicit cache hit | $0.075 | $0.071 | $0.25 | −70% |
| Explicit cache read | $0.075 | $0.071 | $0.17 | −56% |
| Explicit cache write, 5 minutes | $1.875 | $1.767 | $2.50 | −25% |
| Explicit cache write, 1 hour | $3.00 | $2.827 | not listed | n/a |
QwenCloud pricing documentation and model pages, read 18 November 2026. Global deployment is listed in Frankfurt, Hong Kong, Tokyo and US (Virginia).
The Qwen blog says Alibaba lowered prices “with particularly large reductions in cached-input pricing”, and the table bears that out: the cache-hit rate falls from 12.5 per cent of the input price to 5 per cent. That puts Qwen4-Max’s cache ratio level with OpenAI’s newest models, which also bill cache reads at 5 per cent of input.
Three billing rules from the pricing documentation affect planning:
- Batch is half price, but not with caching. The documentation says Batch API input and output rates are 50 per cent of real-time pricing, available in Singapore and Beijing, and that “batch and cache discounts cannot be combined on the same request.” The Qwen4-Max model page lists batch as a feature but, unlike the Qwen3.8-Max page, does not show a separate batch price.
- No long-context surcharge. Qwen4-Max has no input-length tiers, so a 900,000-token request costs the same per token as a 9,000-token one.
- Tools cost extra. Built-in tools such as web search carry per-call fees that differ by region.
A worked agent-loop comparison
The following example is hypothetical; the workload is GenAI Brief’s assumption and only the prices come from the vendors’ published lists. We use the same agent turn as in our GPT-6.2 Astra and Sol pricing analysis: 100,000 tokens of history that hit the cache, 8,000 new input tokens and 4,000 output tokens. For Qwen we assume implicit caching, where QwenCloud lists a cache-hit price but no write price, so new tokens are billed at the ordinary input rate. OpenAI bills new cached tokens at a write rate of 1.25 times input, which we include.
| Per turn (hypothetical) | Cached input | New input | Output | Total |
|---|---|---|---|---|
| Qwen4-Max | $0.0075 | $0.012 | $0.018 | $0.0375 |
| Qwen3.8-Max | $0.025 | $0.016 | $0.024 | $0.065 |
| GPT-6.2 Sol | $0.010 | $0.020 | $0.040 | $0.070 |
| GPT-6.2 Astra | $0.040 | $0.080 | $0.160 | $0.280 |
GenAI Brief calculation from QwenCloud and OpenAI list prices on 18 November 2026. OpenAI’s new-input column is billed at its 30-minute cache-write rate.
At equal token counts, Qwen4-Max costs 42 per cent less per turn than Qwen3.8-Max and about half as much as GPT-6.2 Sol. Equal token counts is the assumption that independent measurement does not support. Artificial Analysis reports that Qwen4-Max generated 127 million output tokens to run its Intelligence Index, against a median of 65.4 million, and puts its average cost at $3.27 per task. GPT-6.2 Sol, at a higher list price, cost $1.01 per task on the same index while scoring higher. A team comparing the two on price should therefore measure tokens per completed task on its own work before trusting the per-token comparison.
Open weights: what is and is not available
This is the question the release most clearly answers. Hugging Face listed no Qwen4 models under the Qwen organisation when we checked on 18 November 2026. Neither Alibaba document links to weights or a licence for Qwen4-Max. The Qwen blog’s closing section says: “Smaller models in the family are coming to open source, built on the same lineage and trained with the same generation of techniques, tuned for what small models do best.” The New Stack, which reported the release under the headline that Alibaba “puts the open source jacket back on the hanger”, says Alibaba “has made no commitment to publish” Max weights.
The one open-source artefact released with Qwen4-Max is tooling, not a model. Alibaba says it is open-sourcing the Qwen-Live-Delivery-Plugin, which supports a “Live Delivery” pattern in which a long-running agent delivers “a usable first version early, then expands and improves it through successive releases.” The GitHub repository was created on 14 November 2026 and carries the Apache 2.0 licence.
What the previous open Max licence required
For teams deciding whether to keep self-hosting Qwen3.8-Max or move to the hosted Qwen4-Max, the earlier licence is the benchmark. The Qwen3.8-2.4T-A95B licence is permissive with two conditions:
- A licensee whose commercial product or service has more than 100 million monthly active users, or more than US$20 million in monthly revenue, must display the model name prominently in the user interface.
- A licensee that runs a “Model as a Service or AI Work Assistant business” with aggregate revenue above US$50 million in any consecutive 12 months must obtain a separate licence before commercial use. Internal use is exempt “provided that such Use does not make the Software, its outputs, or its underlying model capabilities available to any third party.”
Those conditions remain in force for Qwen3.8-Max. Qwen4-Max has no equivalent document because there is nothing to license; its terms are QwenCloud’s service terms.
| Question | Qwen3.8-Max (August 2026) | Qwen4-Max (November 2026) |
|---|---|---|
| Weights downloadable | Yes, Qwen3.8-2.4T-A95B on Hugging Face | No |
| Licence | Custom, with revenue and MAU conditions | None published; API terms apply |
| Active parameters | about 95B | 71B |
| Self-hosting | Possible, multi-node | Not possible |
| Fine-tuning | On your own hardware | Listed as a QwenCloud feature |
| Data location | Your choice | QwenCloud regions |
The trade-off for a self-hosting team is sharp. Qwen4-Max is cheaper per token than Qwen3.8-Max on Alibaba’s API, but a team that chose Qwen3.8-Max to keep data and inference in-house cannot follow it to Qwen4-Max. For that team, the relevant releases are the promised smaller Qwen4 open models, whose sizes and licences are unknown. Our October ranking of frontier models listed Qwen3.8-Max among its open-weight picks; that framing does not carry to its successor.
The benchmark claims, and the independent checks
Alibaba’s claims fall into three groups: its own Work360 benchmark, a comparison chart of public benchmarks, and ranking claims.
Work360 and the 62 per cent figure
Work360 is Alibaba’s framework for professional tasks: “26 Tier-1 domains, 182 Tier-2 subdomains, and 721 Tier-3 specialized fields.” The blog says Qwen4-Max scored 52.69 against 47.74 for Qwen3.8-Max. On token use, it reports reductions of “approximately 24% and 35% at Xhigh and High effort, respectively” on Work360, and “approximately 48% and 49%” on Terminal-Bench 4.0. Combined with the new prices, it says, “the average cost per task drops from $4.43 to $1.67”, a 62 per cent reduction. The press release rounds this to “around 60%.”
The 62 per cent figure therefore has three ingredients, each from Alibaba: a benchmark Alibaba designed, token counts Alibaba measured, and the price change. Only the price change is independently checkable. Analysts quoted by Computerworld made the same point. Sanchit Vir Gogia of Greyhound Research said: “Alibaba has published the mileage it got on its own road, and a buyer’s bill depends on the road the buyer drives.”
How the 62 per cent decomposes
The arithmetic shows how much of the claim rests on each ingredient. A 25 per cent cut in input and output prices, on its own, multiplies the cost of a fixed workload by 0.75. If output tokens also fall by 35 per cent, the token-use reduction Alibaba gives for High effort on Work360 (it does not say which token types that covers), the output side of the bill is multiplied by 0.75 × 0.65, or about 0.49, a 51 per cent saving. The remaining distance to 62 per cent has to come from the input side: the 70 per cent cut in cache-hit pricing and any reduction in input tokens across multi-turn tasks. Alibaba does not publish the input and output split behind the $4.43 and $1.67 figures, so this decomposition is our inference from its stated numbers, not a reconstruction of its calculation.
The practical reading is that roughly half of the claimed saving would survive on any workload with similar token counts, because it comes from the price list. The other half depends on the model using fewer tokens on your tasks, which is the part independent measurement has so far not confirmed.
Token use by effort level
The blog’s second chart plots score against tokens per task at three effort levels, Low, High and Xhigh, for both models. Read approximately from that chart, Qwen4-Max at High effort scores slightly above Qwen3.8-Max at Xhigh on Work360 while using roughly 40 per cent fewer tokens per task, and on Terminal-Bench 4.0 Qwen4-Max’s whole curve sits to the left of Qwen3.8-Max’s, at roughly half the tokens for similar or better scores. The largest jump on Terminal-Bench 4.0 is between Qwen4-Max’s High and Xhigh settings, where the score rises from about 31 to 48.5 for a modest increase in tokens.
Two implications follow if those vendor curves hold. First, a team moving from Qwen3.8-Max at Xhigh could start Qwen4-Max at High and expect similar quality at lower cost. Second, terminal-style agent work appears to need Xhigh, which uses more tokens per task than High, although Alibaba reports a similar saving against Qwen3.8-Max at both settings on Terminal-Bench 4.0 (about 48 and 49 per cent). Both are hypotheses to test; the chart is Alibaba’s, on Alibaba’s harness.
The comparison chart
The Qwen blog’s main chart compares Qwen4-Max with Qwen3.8-Max-0902, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1 on 16 benchmarks. We read the following values from that chart; they are Alibaba’s figures, and footnotes say some, such as Agents’ Last Exam, Arena and Artificial Analysis scores, were cited from those organisations’ leaderboards.
| Benchmark (Alibaba’s chart) | Qwen4-Max | Qwen3.8-Max | Opus 5.5 | GPT-6 Astra | Kimi K3.1 |
|---|---|---|---|---|---|
| DeepSWE v1.1 | 75.7 | 69.3 | 74.2 | 74.1 | 75.5 |
| Terminal-Bench 4.0 | 48.5 | 38.9 | 59.6 | 59.1 | 67.2 |
| ProgramBench | 31.5 | 28.0 | 65.0 | 55.0 | 25.0 |
| Code Arena (Elo) | 1750 | 1676 | 1816 | 1763 | 1686 |
| Agents’ Last Exam v1.1 | 46.7 | 27.9 | 43.4 | 42.6 | 43.4 |
| AA-Briefcase v1.1 (Elo) | 1686 | 1605 | 1801 | 1569 | 1674 |
| HLAB-Vals-1.1 | 21.3 | 15.4 | 9.2 | 7.1 | 11.7 |
| Humanity’s Last Exam | 51.7 | 43.9 | 59.8 | 53.6 | 51.1 |
| HealthBench Professional | 60.0 | 39.6 | 59.9 | 66.0 | 57.2 |
Read by GenAI Brief from the benchmark chart in Alibaba’s Qwen4-Max blog post. GPT-6.1 Sol columns omitted for space.
The chart is more mixed than the press release’s description of “top rankings across leading benchmarks”. Qwen4-Max leads its comparison set on Agents’ Last Exam and the HLAB-Vals legal-agent benchmark, and narrowly on DeepSWE, and trails Opus 5.5, GPT-6 Astra and Kimi K3.1 by large margins on Terminal-Bench 4.0, and Opus 5.5 by more than 30 points on ProgramBench. The comparison set also predates the newest releases: it does not include Claude Fable 5.2, Gemini 4 Carbon or GPT-6.2.
What independent sources show
Two independent sources had Qwen4-Max results when we checked on 18 November 2026.
Artificial Analysis scores Qwen4-Max at 43 on its Intelligence Index v5.0.1, which it describes as “below average among comparable models (median: 43)”. It reports $3.27 per task, 127 million output tokens to run the index, which it calls “very verbose” against a median of 65.4 million, and 31.9 minutes per task against a median of 11.3. Its summary calls the model “below average in intelligence, but well priced when comparing to other models of similar price.” For reference, the same site scores GPT-6.2 Sol at 57 for $1.01 per task.
The Agents’ Last Exam leaderboard, updated 16 November, lists Qwen4-Max in the Codex harness at XHigh effort second by pass rate at 46.7 per cent, behind Claude Fable 5.2 at max on 46.9 per cent, with a partial-credit score of 74.4 per cent. That matches Alibaba’s chart. The leaderboard also records 13.6 million output tokens and an estimated $1,201.45 for the full run, more than GPT-6 Astra’s $1,097.06 despite Qwen4-Max’s much lower list price, because GPT-6 Astra used 5.1 million output tokens. Both independent sources point the same way: Qwen4-Max is competitive on some agent work, and it spends many tokens to get there.
Alibaba’s press release also says Qwen4-Max “ranked 6th on Code Arena.” Its own chart puts it at 1750 Elo, below Opus 5.5 and GPT-6 Astra.
Where it sits among hosted options
Putting the independent figures side by side shows the trade Qwen4-Max offers. All of the following come from Artificial Analysis’s model pages as read on 18 November 2026, at each model’s highest tested effort.
| Model | AA Intelligence Index | Cost per index task | Minutes per task | List price in/out |
|---|---|---|---|---|
| GPT-6.2 Sol (max) | 57 | $1.01 | 7.9 | $2 / $10 |
| GPT-6 Astra (max) | 57 | $6.24 | 20.4 | $10 / $50 |
| Qwen4-Max | 43 | $3.27 | 31.9 | $1.50 / $4.50 |
Measured by Artificial Analysis; GenAI Brief ran no tests.
On this mix of nine evaluations, Qwen4-Max has the lowest list price and neither the lowest cost per task nor the highest score. That does not make it the wrong choice for every workload. Its Agents’ Last Exam result shows it can match much more expensive models on long professional tasks, and teams with data-residency or vendor-diversity requirements may weigh QwenCloud’s regions differently. It does mean that the headline per-token price is a weak guide to the bill.
The cybersecurity and safety claims
Alibaba’s safety section makes three kinds of claim, all based on internal work.
Secure coding. Alibaba says it added code-security checks alongside functionality checks in software-engineering training. On Last Secure Code Bench’s “code reconstruction without security hint” and dual-verification setting, the share of functions passing both functionality and security tests rose “from 22.0% for the previous generation (Qwen3.8-Max) to 31.8%”, and the share secure at function, file and repository level together rose from 8 per cent to 14.7 per cent. Even on Alibaba’s numbers, more than two-thirds of functions in that test did not pass both checks, so generated code still needs review and scanning.
Training-data filtering for cyber. The blog says: “For cybersecurity, we filter data according to the specific purpose of each task, retaining data for defensive tasks such as vulnerability verification, detection, and remediation, while removing data for offensive tasks aimed at enhancing attack capabilities.” It reports that internal security experts reviewing long-horizon trajectories, including full chain of thought, found defence capabilities “improved substantially” and that “no capabilities posing large-scale cybersecurity risks have been observed.”
CBRN. Alibaba says it built dedicated safety data and post-training, and that internal experts found the model’s proposed solutions “are not genuinely innovative and cannot complete the task end-to-end.”
The evidence behind these claims is narrower than the claims. Alibaba publishes no external red-team results, no government evaluation and no benchmark scores for offensive-capability tests. Yash Kadakia of Security Brigade, quoted by Computerworld, questioned whether filtering can separate the two: “offensive and defensive knowledge are largely the same knowledge.” Gogia recommended treating the filter “as a promise until a government evaluator or an established outside red team has tested it.”
Alibaba’s approach also differs from OpenAI’s for GPT-6.2, which serves the same model to everyone and plans to gate more permissive cyber behaviour through vetted access programmes. Alibaba describes removing offensive data at training time. For a defender, a model trained without offensive material may be less useful for exploit validation; for an application owner, the absence of a published external evaluation means the safeguard’s strength is unknown either way.
What remains unknown
Several facts a buyer would want were not in the documents we read on 18 November 2026.
The open-weight family. Alibaba has not named the smaller Qwen4 models, their sizes, their release dates or their licences. Whether they will carry Qwen3.8-Max’s revenue and user thresholds, or a permissive licence such as Apache 2.0, is open.
Batch pricing and caching detail. The model page lists batch processing but no batch price, and the general rule of 50 per cent off cannot be combined with caching. QwenCloud lists no write price for implicit caching; teams should confirm on their own invoices how newly cached tokens are billed.
The parameter count. Alibaba’s two documents describe 2.4 trillion as the backbone and as the total. Without weights or a technical report, the difference cannot be checked.
External safety evaluation. No outside red team or government evaluator has published results on Qwen4-Max’s cyber or CBRN safeguards.
Independent per-task results on real work. Artificial Analysis and Agents’ Last Exam are the only independent sources so far. Neither measures the Work360 tasks behind the 62 per cent claim.
What builders should do
- If you self-host Qwen3.8-Max for data control, stay put for now. Qwen4-Max cannot be self-hosted. Watch for the smaller open Qwen4 models and read their licences when they appear.
- If you use Qwen3.8-Max through the API, test the switch. List prices fall 25 per cent and cache hits 70 per cent, and the thinking-mode output ceiling doubles. Run your own tasks at High and Xhigh and record output tokens per completed task.
- Compare cost per completed task, not per token. Artificial Analysis measured about twice the median output volume; on its index Qwen4-Max cost more per task than GPT-6.2 Sol. Your workload may differ, which is why it has to be measured.
- Use caching deliberately. At $0.075 per million cached tokens against $1.50 uncached, stable prefixes are worth engineering for. Remember that batch and cache discounts do not stack.
- Plan for latency. At 31.9 minutes per Artificial Analysis index task, Qwen4-Max suits background agents better than interactive chat.
- Treat the cyber safeguards as unverified. Keep your own code scanning and tool permissions, and do not rely on the model refusing offensive tasks.
- Check data terms by region. Token Plan is Singapore-only, and Team Edition’s no-training guarantee is specific to that edition; confirm the terms for the API route you use.
Qwen4-Max is a cheaper hosted model with a credible architecture story and a mixed benchmark record. The bigger change is strategic: Alibaba has returned its flagship to the API, and for teams that valued Qwen for its open weights, the useful Qwen4 release has not happened yet.
Sources
- Qwen4-Max: Your New Teammate Has Arrived Qwen Team, Alibaba
- Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery Alibaba Cloud
- Qwen4-Max model page QwenCloud
- Qwen3.8-Max model page QwenCloud
- Pricing QwenCloud documentation
- Token Plan overview QwenCloud documentation
- Qwen-Live-Delivery-Plugin QwenLM on GitHub
- Qwen3.8-2.4T-A95B licence Qwen Team, via Hugging Face
- Qwen4 Max model analysis Artificial Analysis
- Agents' Last Exam leaderboard (ALE v1.1) Agents' Last Exam
- Qwen4-Max now available on AI Gateway Vercel
- Alibaba launches Qwen4-Max with training, attention optimizations SiliconANGLE
- With Qwen4-Max, Alibaba puts the open source jacket back on the hanger The New Stack
- Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards Computerworld
Frequently asked questions
How much does Qwen4-Max cost?
In QwenCloud's Singapore (international) region, $1.50 per million input tokens, $4.50 per million output tokens including thinking, and $0.075 per million cached input tokens. Explicit cache writes cost $1.875 per million for five minutes and $3 for one hour. Beijing and Global deployments list slightly lower rates, such as $1.413 input and $4.24 output.
Are Qwen4-Max's weights open?
No. As of 18 November 2026 Alibaba offers Qwen4-Max only through QwenCloud, its own products and third-party API providers. It has published no weights or licence for it. Alibaba says smaller Qwen4 models will be released with open weights but has not said which or when.
How big is Qwen4-Max?
Alibaba's Qwen blog describes a 2.4-trillion-parameter backbone with 71 billion parameters activated per token, plus an additional 328 billion N-gram Embedding parameters. Alibaba Cloud's press release calls the 2.4 trillion figure 'total parameters', so the two documents count differently.
What is Qwen4-Max's context window?
1 million tokens. QwenCloud lists a maximum input of 991,000 tokens (983,000 in thinking mode), maximum output of 131,000 tokens (262,000 in thinking mode) and a maximum reasoning length of 262,000 tokens. Artificial Analysis lists a 984,000-token context window.
Is Qwen4-Max better than GPT-6.2 Sol or Claude Opus 5.5?
Not on independent aggregate measurement so far. Artificial Analysis scores Qwen4-Max at 43, below GPT-6.2 Sol's 57. Alibaba's own chart shows Qwen4-Max ahead of Opus 5.5 on some agent benchmarks, such as DeepSWE and Agents' Last Exam, and behind it on others, such as Terminal-Bench 4.0 and Code Arena.
What cybersecurity safeguards does Qwen4-Max have?
Alibaba says it filtered training data by purpose, keeping defensive material such as vulnerability detection and removing material aimed at offensive capability, and added code-security checks to software-engineering training. Its evidence is an internal expert review and a secure-coding benchmark score; it has not published external red-team or government evaluation results.


