Alibaba
Qwen4-Max
Alibaba launched Qwen4-Max on 2026-11-16 under API id qwen4-max, the first model in the Qwen4 family, with text, image, and video in, text out, a 1,000,000-token context, and Singapore pay-as-you-go of $1.50/$4.50 per million tokens. It is proprietary. No Qwen4-Max score is stored. Alibaba has not retired Qwen3.8-Max or Qwen3.8-Flash-Next.
Qwen4-Max benchmarks and pricing, every number sourced: no Qwen4-Max benchmark scores tracked yet, priced at $1.50 per million input tokens and $4.50 per million output.
Qwen4-Max architecture: Sparse MoE with native vision; 2.4T backbone total parameters (71B activated per token); 1M-token context window.
- Released
- 2026-11-16
- License
- proprietary
- Context window
- 1M tokensvendor-stated (docs.qwencloud.com/developer-guides/getting-started/latest-model: context length 1,000,000 tokens; checked 2026-11-16)
- Knowledge cutoff
- Not disclosed
- Succeeded
- Qwen3.8-Max (2026-08-03)
- Parameters
- 2.4T backbone (71B active)
- Architecture
- Sparse MoE with native vision
Qwen4-Max API pricing
$1.50 in / $4.50 out per 1M tokens — official pricing source
What Qwen4-Max costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.195 |
| A codebase review | 1,000K / 100K | $1.95 |
| A day of agent work | 10,000K / 1,000K | $19.50 |
Computed from Qwen4-Max’s list rates above — cache discounts and batch tiers are not applied.
Qwen4-Max is one of 3 Alibaba models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| Qwen4-Max | $1.50 | $4.50 |
| Qwen3.8-Flash-Next | — | — |
| Qwen3.8-Max(previous version) | $2.00 | $6.00 |
Notes on the record
Alibaba's @Alibaba_Qwen post is timestamped 2026-11-16T02:01:55Z. Five thread follow-ups are cited by timestamp only: 2026-11-16T02:01:56Z, 02:01:57Z, 02:01:58Z, 02:01:59Z, and 02:02:00Z. Later Arena quote-posts are not launch-spec sources. The post describes a new MoE architecture with native vision and a 1M-token context window. Qwen Cloud docs checked 2026-11-16 call this the first model in the Qwen4 family, under API id qwen4-max. That id is separate from qwen4-max-preview, which those docs still say has its own preview benefits. Pay-as-you-go is live on QwenCloud. Token Plan remains available. The launch post also names Qwen Studio, QwenWork, and Qoder.
The launch blog states a 2.4T-parameter backbone, 71B activated per token, and an additional 328B n-gram embedding parameters. The stored total is that backbone figure. The 328B embeddings sit outside it. The same post says smaller Qwen4 models are coming to open source, with no sizes and no dates. Qwen4-Max itself is proprietary.
Docs checked 2026-11-16 state a context length of 1,000,000 tokens. Maximum output is 131,072 tokens with thinking off. With thinking on, the trace and the final answer share a 262,144-token maximum, and the default thinking budget is 131,072. Thinking is on by default at reasoning_effort=high. The QwenCloud model page displays Max Output 393K, which does not match those two caps, so this page stores the spec-table figures. Batch inference, fine-tuning, and model deployment are not supported.
Singapore list price is the same for thinking and non-thinking, and has no input-length tier: $1.50 input and $4.50 output per million tokens. Cache write is $1.875 for five minutes and $3 for one hour. Cache read is $0.075. Beijing and Global (Frankfurt, Hong Kong, Tokyo, and US Virginia) are $1.413 input and $4.24 output, with cache write $1.767 and $2.827 and cache read $0.071. The stored headline is the Singapore rate. Singapore's one million free tokens are not stored as a $0 price. No Work360, cost-per-task, or Arena figure is stored as a score.
Compare with
FAQ
What does Qwen4-Max cost in Singapore?
Singapore pay-as-you-go, checked 2026-11-16, is $1.50 input and $4.50 output per million tokens. Thinking and non-thinking use that same rate, and there is no input-length tier inside the 1,000,000-token window. Cache write is $1.875 for five minutes and $3 for one hour. Explicit and implicit cache read are $0.075. The stored headline is this Singapore rate.
What do Beijing and Global regions charge?
China North 2 (Beijing) and Global, which covers Frankfurt, Hong Kong, Tokyo, and US Virginia, list $1.413 input and $4.24 output per million tokens. Cache write is $1.767 for five minutes and $2.827 for one hour. Cache read is $0.071. Thinking and non-thinking match. Those rates are not the stored headline. The model-prices page says free quota applies only in Singapore.
Is the Singapore free quota the listed price?
No. New users in Singapore get one million free tokens on pay-as-you-go, separate from Token Plan credits. Other regions list no free quota. This page stores the Singapore list rate of $1.50/$4.50. The free quota is not stored as $0.
What is the context window and output cap?
The latest-model spec table checked 2026-11-16 states a context length of 1,000,000 tokens, stored here as 1000000 rather than 1048576. Maximum output is 131,072 tokens with thinking off. With thinking on, the thinking trace and the final answer share 262,144 tokens, and the default thinking budget is 131,072. Thinking is on by default at reasoning_effort=high. The QwenCloud model page displays Max Output 393K. That display does not match the spec table, so it is not the stored cap.
What inputs and tools does Qwen4-Max support?
Docs checked 2026-11-16 list text, image, and video in, and text out. Function calling, structured output, web search, PDF, code interpreter, and MCP are supported. Code interpreter requires thinking to be enabled. Batch inference is not supported. Fine-tuning and model deployment are not supported. The model page's Batches and Fine-tuning cards point at general docs. The capability table is the spec used here.
Does Qwen4-Max replace Qwen3.8-Max or Qwen3.8-Flash-Next?
Alibaba has not published a retirement for either model, and this record adds no deprecation. Both stay listed at their own prices. Qwen3.8-Max is an earlier Max in the Qwen line, so the catalog's computed previous-version rule labels it a previous version of Qwen4-Max. That label is about the lineup, not an Alibaba retirement notice. Qwen3.8-Flash-Next is a different tier and is not labeled that way. The 2026-11-11 qwen4-max-preview tracker entry is unchanged.
Has Qwen4-Max been independently benchmarked?
No tracked independent board had a Qwen4-Max row confirmed for the benchmarks this site stores, so this page adds no score. The launch blog's Work360 token and cost-per-task comparisons, and later Arena quote-posts, stay off the tracked boards.
What parameter counts does the launch blog state?
The blog at qwen.ai/blog?id=qwen4-max states a 2.4T-parameter backbone, 71B activated per token, and an additional 328B n-gram embedding parameters. This page stores 2.4T backbone and 71B active. The 328B embeddings are additional and are not inside the 2.4T. The same post says smaller Qwen4 models are coming to open source, with no sizes and no dates. Qwen4-Max is proprietary.
Further reading
- Models with 10M token context windows 2026 — Qwen4-Max is one of the 52 models it compares.
- Self-reported vs. independent AI benchmarks — the self-reported-versus-independent rule every score on Qwen4-Max’s page follows.