What's new

Alibaba made Qwen4-Max generally available on 16 November 2026. It's the first model in the Qwen4 family. Alibaba's changelog shows that a restricted preview, qwen4-max-preview, opened five days earlier on 11 November for Token Plan subscribers only, as CASRAI reports. You can now call the model through the QwenCloud API, and it also runs inside Alibaba's QwenWork, Qoder and QwenStudio products.

There are no weights to download. This matters because Qwen3.8-Max, released in August, was the first Max model Alibaba opened up, according to the South China Morning Post. Alibaba says smaller open-weight Qwen4 models are coming but hasn't given a date. The SCMP, which Alibaba owns, also reports the company's plan to scale Qwen 5 to between 5 and 10 trillion parameters. For now, teams that want to run the Max tier on their own hardware are stuck on the previous generation.

How it works

Qwen4-Max is a mixture-of-experts model. Its parameters are split into many expert sub-networks, and a router sends each token through only a few of them. That keeps per-token compute far below what the headline size suggests. The model has 2.4 trillion parameters in total, the same as Qwen3.8-Max, and it uses 71 billion of them for each token. Its predecessor used 95 billion, so that's a 25 per cent cut, and Alibaba says this is the main reason serving costs fell.

The Decoder sets out the rest of the design as Alibaba describes it:

  • Hybrid attention. Three layers squeeze earlier context into a compact state, and a fourth decides which parts of it matter most. This is meant to keep long inputs affordable.
  • Gated layer connections. Four parallel paths link the layers, and gates that adapt to each input control how much information flows along each path.
  • N-gram embeddings. A 328-billion-parameter embedding layer counts toward the 2.4 trillion total.
  • Stable training. Alibaba says the full-scale run had no loss spikes. On its own bits-per-byte tests, the base model scored 5 to 13 per cent better than Qwen3.8-Max-Base.

Alibaba hasn't released a technical report, so nobody outside the company has checked any of these claims.

Limits and features

The QwenCloud model page lists the model ID as qwen4-max. It accepts text, images and video and returns text. The context window is 1M tokens. Without thinking mode, input can run to 991K tokens and output to 131K. With thinking turned on, input drops to 983K and the output ceiling, which covers both reasoning and the answer, rises to 262K. Each account is limited to 5 million tokens per minute.

The page also lists built-in tools: web search, web extraction, a code interpreter, and image search in both directions through the Responses API, plus web fetch and web search through the Anthropic-compatible API. CASRAI's summary of the developer guide adds function calling, structured output, PDF understanding and explicit prefix caching. The same guide says batch inference, fine-tuning and dedicated deployment aren't available. CASRAI also notes that the listing and the guide disagree on some smaller labels, so check each feature you need in your own account before you rely on it.

What it costs

These are QwenCloud's rates in US dollars per million tokens:

  • Standard input: $1.50
  • Output: $4.50
  • Cache reads, implicit or explicit: $0.075
  • Explicit cache writes: $1.875 for a 5-minute cache, $3.00 for a 1-hour cache

MLQ describes these as international list prices in Singapore. Qwen3.8-Max cost $2.00 for input and $6.00 for output. The Decoder puts the cut at 25 per cent on standard tokens and 70 per cent on cache hits. The SCMP's figure of "about 8 US cents" for input lines up with the cache-hit rate. The standard input price is $1.50. CASRAI warns that Beijing and Global deployments have their own price lists, and that pay-as-you-go and Token Plan accounts need different keys and endpoints. So when you compare quotes, check that the region, plan and billing category match.

Caching makes the biggest difference on long prompts. We worked out an example from the listed rates. A call with 200K input tokens and 10K output tokens costs about US$0.35. If 180K of that input is a repeated prefix that hits the cache, the same call costs about US$0.09. If your agent resends a big system prompt or a whole codebase on every turn, keep that block byte-identical at the start of the prompt so it caches.

How far to trust the benchmarks

Alibaba's headline claim of "about 60 per cent cheaper per task" comes from Work360, a benchmark it built itself. MLQ says it spans 26 work domains and 721 specialist fields. On Work360, the score rose from 47.74 to 52.69 and the average task cost fell from US$4.43 to US$1.67, a 62 per cent drop. Nobody else has reproduced these numbers.

The Decoder reports figures from Artificial Analysis, which come from outside Alibaba. On that index, Qwen4-Max scores 43, up from 34 for Qwen3.8-Max. Among Chinese models, it sits second behind Kimi K3.1 Pro on 46. Running the index cost US$3.27 per task, against US$10.02 for the old model and US$3.68 for Kimi. Average output fell from about 171,000 tokens per task to 105,000. We haven't seen these scores on Artificial Analysis's own site, so treat them as second-hand.

The model has clear weak spots. On Terminal-Bench 4.0, which tests command-line agents, it scores 48 per cent to Kimi K3.1 Pro's 67. It also trails Claude Opus 5.5 and the GPT models on Humanity's Last Exam. The efficiency gains look more solid. MLQ reports about 35 per cent fewer tokens on Work360 at the High reasoning setting, and 48 to 49 per cent fewer on Terminal-Bench 4.0. Shorter answers lower both your bill and your wait, even before you count the price cut.

Safety changes

According to the SCMP, Alibaba stripped offensive-hacking material out of the training data. It kept defensive material such as vulnerability detection, and it added post-training data aimed at chemical, biological, radiological and nuclear risks. On a secure-coding benchmark, the share of functions that passed both functional and security checks rose from 22.0 to 31.8 per cent. Alibaba's own staff did the risk reviews. If you build security tools, expect some red-team prompts to be refused, and test for that early.

Move off the preview

The preview endpoint shuts down at midnight on 25 November 2026, UTC+8, according to AI TLDR's summary of Alibaba's preview terms. That's 5am on Wednesday 26 November in Auckland. If any coding tool still points at qwen4-max-preview or the Token Plan preview base URL, switch it to qwen4-max before then.

What builders can do with it

Alibaba also published the Qwen Live Delivery Plugin under Apache-2.0. It makes a long-running agent ship a small working version first and then improve it in later rounds. You watch progress in a local browser workbench, where you can add requirements, change the time budget or stop the task. Version 0.1.0 only supports OpenAI's Codex CLI on Linux and macOS, needs Node.js 22 or later, and doesn't include a model or API key. You install it with ./live-delivery/install.sh --workspace /path/to/project, set a budget with a variable such as QWEN_LD_TIME_BUDGET=12m, and start a task with $ld:go. Keep the default workspace-write sandbox. Setting QWEN_LD_TASK_SANDBOX=danger-full-access removes both the sandbox and per-command approval. Keep the workbench link private, too.

For teams in Aotearoa, QwenCloud's sample code uses the ap-southeast-1 (Singapore) endpoint, which is the closest one we found. Before you send client or personal data, confirm where prompts are processed and stored. The 1M-token window suits agents that work across a whole repository and long document reviews. CASRAI points out that a big window doesn't mean the model will find every relevant passage, so test retrieval on your own files first. If you build something with it, share it with members through ASAI projects.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member