Qwen4-Max is Alibaba's new flagship, and this time it's closed
Qwen4-Max scores 43 on Artificial Analysis, below open-weight Kimi K3.1 Pro at 46, and Alibaba kept its weights closed this time.
Senior Full-Stack Engineer- Published
- 8 minReading time
- Artificial IntelligenceCategory
Listen to this post· AI-narrated
0:00 / 10:06Alibaba released Qwen4-Max on 16 November, the first model of its Qwen4 generation, and one of the first replies on Hacker News (opens in a new tab) went straight to the independent chart: "Artificial Analysis score doesn't look too hot. Doesn't even surpass Kimi 3.1". The score is only part of the launch. Qwen4-Max runs on a new architecture that activates a quarter fewer parameters than Qwen3.8-Max, reads images and video natively, holds a million tokens of context and costs 25% less per token. And it ships without weights. In August, Qwen3.8 put a Max-class Qwen model in the open for the first time. Three months later, Alibaba kept Qwen4-Max closed and promised only smaller open Qwen4 models, with no date.
TL;DR: Qwen4-Max brings a new architecture, 2.4T parameters with 71B active and native vision, at $1.50 per million input tokens. It scores 43 on the Artificial Analysis Intelligence Index, below the open-weight Kimi K3.1 Pro at 46, and it ships API-only, unlike Qwen3.8's Max-class model. Kimi stays the open model to beat.
What's new in Qwen4-Max?
Qwen4-Max is built on a new Qwen4 architecture with a 2.4-trillion-parameter backbone, of which 71 billion are active for each token, plus 328 billion parameters in what Alibaba calls N-gram Embedding. Qwen3.8-Max was also 2.4T, with 95B active, so the new model does its work with about a quarter fewer active parameters. Fewer active parameters per token make a big model cheaper to serve.
The change that matters most for developers is how it reads long inputs. According to the launch post (opens in a new tab), for every sparse attention layer, three Gated DeltaNet layers squeeze everything read so far into a fixed-size state, and a small indexer picks which parts of the input the full attention layer looks at. Alibaba says this cuts attention and indexing costs on long sequences. The model docs (opens in a new tab) back the result with a 1,000,000-token context, image and video input in the same model, and up to 262,144 tokens of thinking and answer per call.
Two smaller changes round it out: four gated residual branches, which Alibaba credits for a 2.4T pre-training run with no loss spikes, and N-gram Embedding, a lookup of short token patterns that adds capacity for little compute. On Alibaba's in-house tests, the base model predicts text 5 to 13% better than Qwen3.8-Max-Base, measured in bits per byte, while activating 25% fewer parameters. Those are Alibaba's numbers on its own checkpoints, and the long-context design is the piece the smaller open Qwen4 models are meant to inherit.
How do you use Qwen4-Max from a coding agent?
Qwen4-Max speaks the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages APIs, so Claude Code, Codex and OpenCode can use it by changing the base URL; QwenCloud, Alibaba's API platform, has a setup page for each. OpenCode (opens in a new tab) added it on launch day.
Alibaba claims the model also gets more done per token. On Work360, its own benchmark of industry tasks, token use fell by 24 to 35% and the average cost per task dropped from $4.43 to $1.67, a 62% reduction against Qwen3.8-Max. For long jobs it adds Live Delivery: the agent hands over a working first version early and updates it as it goes, so you can redirect it mid-task. Alibaba released the Live Delivery plugin (opens in a new tab) as open source.
Four settings decide what you pay:
- Reasoning effort comes in three levels,
low,highandxhigh, and defaults tohigh. The preview that has run since 11 November on Token Plan, Alibaba's subscription for coding tools, defaults toxhigh, so moving from the preview toqwen4-maxchanges behaviour even with the same prompt. - The output cap is shared.
max_completion_tokenslimits thinking and the final answer together, so a long think leaves less room for the answer. - Explicit caching is new for Qwen4, with 5-minute and 1-hour lifetimes. Cached input reads cost $0.075 per million tokens, down from $0.25 on Qwen3.8-Max, 70% less.
- Price. $1.50 per million input tokens and $4.50 per million output, thinking included, down from $2 and $6 for Qwen3.8-Max.
How good is Qwen4-Max against Opus 5.5 and Kimi K3.1 Pro?
Qwen4-Max scores 43 on the Artificial Analysis Intelligence Index (opens in a new tab), which runs nine evaluations on Artificial Analysis's own harness, Terminal-Bench 4.0 (real terminal tasks) among them. That is above the median of 41 for models in its price range and well short of the top:
| Model | Weights | Intelligence Index | Cost per index task |
|---|---|---|---|
| Claude Opus 5.5 (max) | closed | 59 | $10.04 |
| Kimi K3.1 Pro (Moonshot AI) | open | 46 | $3.68 |
| Qwen4-Max | closed | 43 | $3.27 |
Kimi K3.1 Pro scores three points higher for 41 cents more per task. Opus 5.5 costs three times as much per task and scores 16 points more. Qwen4-Max's cheap tokens also go partly to volume: it wrote 127 million output tokens to run the index, against a median of 78.6 million for its price range.
Alibaba's own chart doesn't change the order. Of the 16 benchmarks it chose, Qwen4-Max leads on 4 and Opus 5.5 on 7. It edges Opus on DeepSWE, an agentic coding test, 75.7 to 74.2. On the harder coding tests it falls behind: 48.5 on Terminal-Bench 4.0 against 59.6 for Opus 5.5 and 67.2 for Kimi K3.1, as the chart labels it, and 31.5 on ProgramBench, which asks a model to rebuild software from its behaviour, against 65.0 for Opus.
What does a closed Qwen4-Max mean for open weights?
Qwen4-Max goes back to how Qwen's Max models worked before August: you can only rent it. In August, Alibaba published Qwen3.8-2.4T-A95B (opens in a new tab) on Hugging Face, and its model card says it plainly: "For the first time, Qwen3.8 brings a Qwen-Max-class model to open release." Qwen3.8-Max is the hosted version built on those weights, and the weights carry a Qwen licence that allows commercial use with restrictions. They were downloaded 28,798 times in the last month.
Qwen4-Max has no repository. The launch post closes with a promise instead: "Smaller models in the family are coming to open source, built on the same lineage". It gives no sizes or dates, and Qwen's Hugging Face page has no Qwen4 model yet.
The difference is what you're allowed to do. With Qwen3.8's weights you could run the flagship on your own hardware or a host you chose, fine-tune it and keep requests in your own region. QwenCloud's docs say Qwen4-Max supports neither fine-tuning nor deployment, and its regions page (opens in a new tab) explains that the Global inference scope, which the Tokyo endpoint uses for this model, can process a request on any node, mainland China included. A team that chose Qwen because the best one could be self-hosted loses that option with this generation.
I'd take this release as two pieces of news. The architecture is new and the price is lower. The open-weight ranking doesn't move: the best model you can download is still Kimi's, and it outscores Qwen4-Max itself.
Which open-weight model is best right now?
Kimi K3.1 Pro is the highest-scoring open-weight model on the Artificial Analysis leaderboard, at 46, three points above the closed Qwen4-Max. Below it, the open field drops off fast:
| Open-weight model | Intelligence Index |
|---|---|
| Kimi K3.1 Pro (Moonshot AI) | 46 |
| DeepSeek V4.1 Flash 1017 (max) | 33 |
| GLM-5.3 (max), Z AI | 32 |
| MiMo-V2.6-Pro (Xiaomi) | 30 |
| Qwen3.8 27B (xhigh) | 20 |
The gap between 46 and the low 30s is where the smaller Qwen4 models will be judged. Unsloth (opens in a new tab), which publishes quantized builds of open models for local use, says Qwen4-27B will run on 16 GB of RAM. The model it follows, Qwen3.8 27B, scores 20. Until Alibaba publishes the 27B and someone scores it, nothing from Qwen challenges Kimi's lead.
What to do this week
If you call models through an API and judge them by cost per finished task, run Qwen4-Max on high and Kimi K3.1 Pro on the same ten tasks before you switch: Artificial Analysis has Kimi ahead by three points for 41 cents more. If you're on the preview, set the reasoning effort yourself so the default doesn't change under you. If you need weights you can host, Kimi K3.1 Pro is the strongest you can download today, and Qwen3.8-2.4T-A95B is still there if you need a Qwen model. I didn't run these benchmarks myself; every number comes from the sources linked above.
I write about coding agents and the models behind them on the gabs.app blog. I'm a full-stack engineer and my work is on gabs.app; if you're hiring, the contact form comes straight to me.