Qwen4-Max: specs, benchmarks, pricing, open-weights status & API options
Qwen4-Max is Alibaba’s new flagship, generally available since November 16, 2026 after a five-day preview. The independent number first: Artificial Analysis measures it at 43 on its Intelligence Index (v5.0.1) — nine points above Qwen3.8 Max (34) on the same index version, at a 25% lower list price: $1.50 input / $4.50 output per 1M tokens, against $2 / $6 for its predecessor.
It is also, today, a hosted API and nothing else. There is no Qwen4 repository on Alibaba’s Hugging Face organization, no model card and no license. Alibaba’s launch materials say smaller Qwen4 models will be open-sourced later — with no date, no sizes and no license named. That status is the part of this page we will keep current: see Open weights and license.
Qwen4-Max specs
Section titled “Qwen4-Max specs”| Released | November 16, 2026 (general availability). Preview id qwen4-max-preview opened November 11 and retires November 25 |
| Architecture | Sparse Mixture-of-Experts, 2.4T total parameters / 71B active per token (Alibaba’s figures) — a new base model, not a post-trained Qwen3.8 |
| Input / output | Text, image and video in; text out |
| Context window | 1M tokens (991,808 max input; 983,616 with thinking on) |
| Max output | 131,072 tokens; 262,144 with thinking, reasoning included |
| Reasoning | Thinking on by default. Effort levels low, high (default) and xhigh; thinking can be disabled per request |
| API features | Function calling with parallel tool calls, structured output (JSON object and JSON schema), implicit and explicit context caching, web search, prefix completion |
| Compatibility | OpenAI Chat Completions, OpenAI Responses and Anthropic Messages on Alibaba’s API |
| Model id (Alibaba) | qwen4-max |
| List price | $1.50 in / $4.50 out per 1M tokens (international); cached input $0.075 |
| Open weights | Not published. No Qwen4 weights, model card or license as of November 17, 2026 |
| Measured speed | 54 output tok/s (Artificial Analysis) |
Limits and prices from Alibaba’s model page, read on November 17, 2026.
Where it lands on the Intelligence Index
Section titled “Where it lands on the Intelligence Index”Artificial Analysis moved its Intelligence Index to v5.0.1 this month. Scores are not comparable with the v4.3.2 figures quoted across the web (and in our own older posts) until last week: every model was re-scored, most of them downward. All numbers below are v5.0.1, read from each model’s Artificial Analysis page on November 17, 2026.
| Model | AA Index (v5.0.1) | Input $/1M | Output $/1M | Weights |
|---|---|---|---|---|
| Claude Fable 5.2 | 65 | $8.00 | $40.00 | closed |
| Gemini 4 Carbon | 60 | $2.00 | $10.00 | closed |
| Claude Opus 5.5 | 59 | $4.00 | $20.00 | closed |
| GPT-6 Astra | 57 | $10.00 | $50.00 | closed |
| Claude Sonnet 5.5 | 55 | $2.00 | $10.00 | closed |
| GPT-6.1 Terra | 51 | $1.00 | $5.00 | closed |
| Kimi K3.1 Pro | 46 | $2.00 | $10.00 | open |
| Qwen4-Max | 43 | $1.50 | $4.50 | closed |
| Qwen3.8 Max | 34 | $2.00 | $6.00 | open |
Honest framing: Qwen4-Max does not reach the top of the index — it sits 22 points under the leader, Claude Fable 5.2 (65) — and it is three points behind the highest-ranked open-weights model, Kimi K3.1 Pro. What it does is move Alibaba’s flagship nine index points in one generation while cutting the list price by a quarter, which puts it on the price-vs-intelligence Pareto frontier of the models we track: nothing scoring 43 or more lists a lower output price.
Two caveats from the same Artificial Analysis page. Qwen4-Max is verbose: it generated 127M tokens to complete the index, so its measured cost is $3.27 per index task despite the low per-token price. And it is slow per task: 32.2 minutes per index task against a median of 13 across the models Artificial Analysis tracks, at 54 output tok/s. Full field and frontier in our monthly LLM Pareto Frontier report.
The vendor benchmarks
Section titled “The vendor benchmarks”Alibaba’s launch post compares Qwen4-Max with Qwen3.8 Max, Kimi K3.1, Claude Opus 5.5 and the GPT-6 line. The figures below are vendor-reported (Alibaba’s own runs or third-party leaderboards it cites), as relayed by launch-day coverage; we have not reproduced them.
| Benchmark (as reported by Alibaba) | Qwen4-Max | Kimi K3.1 | Claude Opus 5.5 |
|---|---|---|---|
| DeepSWE v1.1 | 75.7 | 75.5 | 74.2 |
| Terminal-Bench 4.0 | 48.5 | 67.2 | 59.6 |
- Coding is a tie at the top, not a win. On DeepSWE v1.1 the three models sit within 1.5 points of each other.
- Terminal work is the weak spot. On Terminal-Bench 4.0 Qwen4-Max trails Kimi K3.1 by almost 19 points. The independent Vals run measured 48.48%, in line with Alibaba’s 48.5.
- Work360, Alibaba’s in-house evaluation across professional fields: 52.69, up from 47.74 for Qwen3.8 Max. Alibaba reports the average cost per task falling from $4.43 to $1.67 — a 62% cut measured against its own previous model, not against competitors.
- Fewer tokens for the same work: about 24% (xhigh effort) to 35% (high effort) fewer on Work360, and about 48% fewer on Terminal-Bench 4.0, versus Qwen3.8 Max — Alibaba’s figures.
Pricing: Qwen4-Max API vs Qwen3.8 Max
Section titled “Pricing: Qwen4-Max API vs Qwen3.8 Max”| Qwen4-Max | Qwen3.8 Max | |
|---|---|---|
| Input, per 1M tokens | $1.50 | $2.00 |
| Output, per 1M tokens | $4.50 | $6.00 |
| Cached input, per 1M tokens | $0.075 | — |
| AA Intelligence Index (v5.0.1) | 43 | 34 |
Reasoning tokens bill as output. Alibaba also offers the model through its subscription Token Plan and inside its own products (Qwen Studio, Qwen Work, Qoder).
Open weights and license: what is and is not published
Section titled “Open weights and license: what is and is not published”As of November 17, 2026:
- Qwen4-Max is closed. It shipped as a hosted API. Alibaba has not announced open weights or a model card for the Max tier.
- No Qwen4 weights exist publicly. The Qwen organization on Hugging Face has no Qwen4 repository; its newest language-model weights are the Qwen3.8 family from August 2026.
- What Alibaba has said: smaller Qwen4 models will be open-sourced later. No date, no parameter sizes and no license have been given. Reports of a confirmed “Qwen4-27B under Apache 2.0” are not backed by any Alibaba announcement we could find.
- No license to read yet. For reference, the Qwen3.8 generation shipped under three different texts: Apache 2.0 for the dense Qwen3.8-27B, a custom Qwen3.8-Max License for the open release of the Max (Qwen3.8-2.4T-A95B), and a community license for Flash-Next. Which of those patterns Qwen4 follows is unknown.
We will update this section the day a Qwen4 repository or license appears.
Where it stands in our review
Section titled “Where it stands in our review”Qwen4-Max is under review, waiting on the licensing gate. Every model in our pools belongs to a generation whose weights are published under a license that permits commercial serving — it is the line we held for Qwen3.8 Max, which joined the Flagship Pool on August 14, the day after its open-weight release. For Qwen4 neither the weights nor the license exist yet, so there is no date from us either.
Today the Flagship Pool serves Kimi K3.1 Pro (46 on the index, the top open-weights score) and Qwen3.8 Max, unlimited during your reserved hours, from $199/mo. The live pipeline status is on the models-under-review page, and additions land in the changelog the day they ship.
Common questions
Section titled “Common questions”What is Qwen4-Max? Alibaba’s flagship model, released on November 16, 2026: a 2.4T-parameter Mixture-of-Experts with 71B active parameters, text, image and video input, a 1M-token context window and reasoning that is on by default. It scores 43 on the Artificial Analysis Intelligence Index (v5.0.1), nine points above Qwen3.8 Max.
How much does the Qwen4-Max API cost? Alibaba lists $1.50 per 1M input tokens and $4.50 per 1M output tokens internationally, with cached input at $0.075 — 25% below Qwen3.8 Max ($2 / $6). Artificial Analysis measures $3.27 per Intelligence-Index task, because the model spends many reasoning tokens.
Is Qwen4-Max open source? Does it have open weights? No. Qwen4-Max is a closed, hosted model, and as of November 17, 2026 no Qwen4 weights, model card or license have been published. Alibaba has said smaller Qwen4 models will be open-sourced later, without a date, sizes or a license.
When will Qwen4 open weights be released, and under which license? Unknown. Alibaba has given no date and named no license. The previous generation used Apache 2.0 for Qwen3.8-27B and a custom license for the open release of the Max, but nothing has been announced for Qwen4.
How does Qwen4-Max compare to Qwen3.8 Max? Nine points higher on the Artificial Analysis Intelligence Index (43 vs 34, v5.0.1) at a 25% lower list price, with video input added and, per Alibaba, 24–35% fewer tokens for the same work on its Work360 benchmark. Qwen3.8 Max has an open-weight release; Qwen4-Max does not.
Is Qwen4-Max better than Kimi K3.1 Pro or Claude Opus 5.5? Not on the independent index (v5.0.1): Claude Opus 5.5 scores 59 and Kimi K3.1 Pro 46, against 43 for Qwen4-Max. On Alibaba’s own chart it edges both on DeepSWE v1.1 (75.7 vs 75.5 and 74.2) and trails both on Terminal-Bench 4.0 (48.5 vs 67.2 and 59.6). Its advantage is price: $4.50 per 1M output tokens against $10 and $20.
Is there an unlimited Qwen4-Max API subscription? Not from us today — Qwen4-Max is under review and waiting on its open-weights license (live status). The Flagship Pool serves Qwen3.8 Max and Kimi K3.1 Pro unlimited from $199/mo: flat monthly fee, no token caps during your reserved hours.
CheapestInference serves Kimi K3.1 Pro and Qwen3.8 Max (Flagship Pool), GLM 5.3 and MiniMax M3 (Frontier Pool) and DeepSeek V4.1 Flash and MiMo V2.6 Flash (Core Pool) through one OpenAI- and Anthropic-compatible API on unlimited time-block subscriptions. See the pools or get started.