Alibaba's top Qwen4 model, taking over from Qwen3.8-Max.
Artificial Analysis scores Qwen4 Max 43 on its Intelligence Index, read on 17 November 2026, at 53 output tokens per second. It suits agent jobs that can wait several minutes and replaces Qwen3.8-Max for teams on Alibaba's API, but it takes 139 seconds to a first answer token with reasoning on.
Checked 17 Nov against Alibaba Cloud's own pages.
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.045 | $0.0045 | $0.050 |
| Support reply | $0.0030 | $0.0013 | $0.0043 |
| One coding agent run | $0.300 | $0.090 | $0.390 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Alibaba Cloud API rates per 1M tokens as Artificial Analysis lists them: $1.50 input, $4.50 output, with cached input 95% cheaper. Qwen says it lowered prices against Qwen3.8-Max.
Qwen4 Max is Alibaba's flagship Qwen4 model, released on 16 November 2026. It is a sparse mixture-of-experts system with 2.4 trillion backbone parameters, 71 billion active per token, a one-million-token context window and native text, image and video input. Its weights are closed, and access runs through Qwen Cloud and OpenRouter.
Reads
Writes
Sparse mixture-of-experts at 2.4T parameters
71B parameters are active per token, plus 328B N-gram embedding parameters, per Qwen's launch post.
One-million-token context window
Qwen states 1M tokens; Artificial Analysis lists 984k.
Text, image and video input
Output is text only, with tool calling and structured outputs listed by OpenRouter.
Adjustable reasoning effort
The effort is chosen per request, or reasoning can be turned off, with the levels documented on OpenRouter.
Live Delivery for long agent runs
The agent ships a usable first version early and improves it in releases; Qwen open-sourced a Live Delivery plugin.
Artificial Analysis Intelligence Index score (v5.0.1)
Source: Artificial Analysis Qwen4 Max page, read 17 November 2026Priced around the middle of the 106 GA models with a published price (rank 60), and rank 47 of 59 on output speed as cited from Artificial Analysis.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $2.25/M | #60 / 106 | peer median $1.73/M |
| Output speedtokens/s | 53 tok/s | #47 / 59 | peer median 94 tok/s |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Alibaba Cloud | 1M | $1.50 / $4.50 | text, image, video → text | |
| Moonshot AI | 1M | $2 / $10 | text, image → text | |
| Z.ai | 1M | $1.40 / $4.40 | text → text | |
| Tencent | 1M | $0.07 / $0.281 | text, image → text |
Training data
Qwen's launch post says cybersecurity training data was filtered by task purpose: data for defensive work such as vulnerability verification, detection and remediation was kept, and data for offensive tasks was removed. It publishes no pre-training data sources or cutoff.
| Name | Endpoint | Auth |
|---|---|---|
| Qwen Cloud | https://www.qwencloud.com/ | API key |
| OpenRouter | https://openrouter.ai/qwen/qwen4-max | API key |
Reasoning is on by default
OpenRouter says reasoning is enabled at high effort unless changed, and Artificial Analysis links that setting to a long wait before the first answer token.
Workaround: Set reasoning effort to low, or turn reasoning off, for short interactive requests.
Two context window figures
Qwen's launch post states a one-million-token window while Artificial Analysis lists 984k tokens.
Workaround: Size prompts that approach the limit against 984k tokens until Qwen publishes an exact figure.
qwen4-max · 16 Nov 2026
Initial release of the Qwen4 flagship; the previous flagship checkpoint was Qwen3.8-Max (0902).
Artificial Analysis lists $1.50 per million input tokens and $4.50 per million output tokens on the Alibaba Cloud API, and OpenRouter shows the same endpoint 5% lower at $1.425 and $4.275. Cache hits are billed 95% below the input rate, and Qwen says its biggest cut against the previous flagship was on cached input. Check Qwen Cloud's own pricing page before a large commitment, since it is the final authority.
No. Artificial Analysis labels Qwen4 Max a proprietary model, and Qwen's launch post says only smaller Qwen4 models are coming to open source. Hokai found no free API tier on the pages it read. Qwen says the model also runs inside QwenWork, Qoder and QwenStudio.
Reasoning is on at high effort by default, which Artificial Analysis links to a very long wait before the first token, plus 53 output tokens per second and 33.2 minutes per Intelligence Index task. OpenRouter lists effort levels from low to xhigh and an off switch, and that is the lever Qwen exposes. Hokai found no measured speed for those settings.
Qwen's launch post reports Work360 at 52.69 against 47.74 and about 48% fewer tokens on Terminal-Bench 4.0 at xhigh effort, both vendor-reported. The independent side is thinner: Artificial Analysis has scored Qwen4 Max, but Hokai found no score for both models on one index version. On vendor figures alone the newer model wins, so wait for a shared independent score if the switch is costly.
Besides Qwen Cloud, OpenRouter resells it from one provider, Alibaba Cloud International, which showed 99.17% uptime over its three-day window on 17 November 2026. The Qwen apps QwenWork, Qoder and QwenStudio also run it, per Qwen's launch post.
Qwen4 Max is the flagship of Alibaba's Qwen4 family, built by Alibaba Cloud and announced on 16 November 2026 as a step up from Qwen3.8-Max. Qwen's launch post describes a sparse mixture-of-experts design with 2.4 trillion backbone parameters, 71 billion activated per token and a further 328 billion N-gram embedding parameters. Attention alternates three Gated DeltaNet layers with one Qwen Sparse Attention layer, and four parallel residual branches use read and write gates. Earlier steps in the lineage are Qwen3.7-Max and the smaller Qwen3.8 27B. A Token Plan-only preview of the model is listed separately as Qwen4-Max-Preview.
Qwen reports its own Work360 benchmark at 52.69 for Qwen4 Max against 47.74 for Qwen3.8-Max. That is vendor-reported, from the 16 November post, and Work360 is Qwen's own taxonomy of 721 specialised work fields, not an independent board. Artificial Analysis, read on 17 November 2026, scores the model 43 on its Intelligence Index v5.0.1 and measures 53 output tokens per second. Across that index run the model wrote 127 million output tokens, which the page calls very verbose, and it needed 33.2 minutes per task. Qwen also says token use on Terminal-Bench 4.0 falls by about 48% at xhigh effort and 49% at high effort against Qwen3.8-Max.
Qwen states a one-million-token context window; Artificial Analysis lists 984k. Reasoning is on by default, and OpenRouter documents effort levels of low, high or xhigh (default high), or reasoning switched off. The launch post gives no maximum output length.
The model accepts text, image and video input and returns text only. OpenRouter lists tool calling and structured outputs as supported. Qwen's post shows it driving Blender, Unity, Unreal Engine and Godot, calling 90 MCP tools in a finance research demo and running a 48-hour closed-loop lab workflow with a partner platform. These are vendor showcases, not independent tests.
Qwen says the model is available on Qwen Cloud through a Token Plan or the pay-per-token API, and inside QwenWork, Qoder and QwenStudio. OpenRouter lists a single provider, Alibaba Cloud International, with no routing. The weights are not released: Artificial Analysis labels the model proprietary, and Qwen says smaller Qwen4 models will come to open source. Teams that need open weights today can look at the previous generation's Qwen3.8 2.4T A95B.
On safety, Qwen reports that the share of code passing both functionality and security checks at function level on Last Secure Code Bench rose from 22.0% for Qwen3.8-Max to 31.8%, and that code secure at function, file and repository level rose from 8% to 14.7%. It says internal experts found no capability posing large-scale cybersecurity risk and no meaningful CBRN uplift. The launch post is the only safety document Hokai found; there is no separate system card.
Who should skip it: anyone building interactive chat, because Artificial Analysis measures 139.46 seconds to the first answer token with reasoning on. Teams that need weights to self-host should compare Kimi K3.1 Pro or GLM-5.3, both open-weights models with 1M-token windows on Hokai, and buyers who want a US vendor contract can weigh Claude Opus 5.5.
Qwen's post gives no training cutoff and no data sources for pre-training. It does say that for cybersecurity it kept data for defensive tasks such as vulnerability verification and removed data aimed at building attack capability. The Qwen site footer lists a Training Data Summary and a Usage Policy.
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Alibaba Cloud's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.