# Qwen 3.7 and 3.7 Max: what the model for coding can do in 2026
> Qwen 3.7 versions and regional prices from Alibaba documentation, vendor-reported comparisons, evidence limits, and a pilot for measuring real task cost.
Source: https://ailoot.ai/blog/en/qwen-3-7-max-coding/
Canonical: https://ailoot.ai/blog/en/qwen-3-7-max-coding/
Language: en
Publisher: Ailoot (https://ailoot.ai/)
Translation: guide-206
Revision: sha256-ea45f6ef6194a978b0bb6d68694fcf8baca58c653f0c0ec32b1372511a1ee59c
Highlights
Qwen 3.7 is a generation of Alibaba Cloud models released between May and July 2026. It includes three models: the top-tier Max, mid-tier Plus, and entry-level Flash. All are closed-source and available only through paid API access; this generation has no open 27B or 35B weights. Since August 3, Qwen 3.8 Max has been the flagship of the lineup.
People come to Qwen 3.7 with two different questions. The first wants to download the weights—the model files—so they can run it on their own computer without the internet or a bill for every request. The second is considering it as a Claude replacement because they have heard it costs three times less. The first will be disappointed right away; the second will have to do the math.
Below is the lineup and regional pricing from the cited Alibaba Cloud documentation, followed by vendor-reported comparisons and a practical test method. Independent rankings need a primary measurement link or a retrievable snapshot before they can support a model choice; no such evidence is linked here.
[Ailoot documentation](https://ailoot.ai/docs/) — reference materials on using the service.
Attention
This material is for reference only and is not legal advice. Check the terms with your bank before paying.
## What is Qwen 3.7, and who is behind it?
Highlights
Qwen is a lineup of models from Alibaba Cloud, Alibaba's cloud division. Qwen 3.7 is the May–July 2026 generation, whose top-tier model is called Qwen 3.7 Max. All models in this generation are closed-source and available only through paid API access. In the terminal, they are used through Alibaba's own tool, called Qwen Code.
Alibaba has been releasing models since 2023 and made their weights openly available through 2026. The Qwen 2.5 generation ranged from 0.5B to 72B, all under a free license, with cost-free access as its main selling point.
By the 3.7 generation, the approach had flipped. The top-tier models were closed: you can no longer download them, only pay for access. The mid-tier of the previous generation remains open. The implications for anyone who wanted to download the model files are covered below in the section on open weights.
Qwen 3.7 Max was announced at Alibaba's cloud conference in Hangzhou on May 20, 2026. It is positioned as an agentic model: long stretches of independent work instead of short chat replies. For a vibe coder, that means the model is aimed squarely at the work you do with Claude Code—project files, terminal commands, and a sequence of steps without your involvement.
## How does 3.7 Max differ from Plus, Flash, and the previous 3.6?
Highlights
There are three models in the 3.7 generation, and their prices differ by dozens of times. Max is the top-tier model, released May 20. Plus is mid-tier, released in late May, and about six times cheaper at Alibaba's list price. Flash is the entry-level and newest model, released in mid-July, and fifty times cheaper than Max. Qwen 3.6 is the previous generation, from April 2026, and its models are the last Qwen models with open weights.
The confusion starts because all versions are sold at the same time under the name “Qwen 3.x.” You can't tell from the version number: Max and Flash differ in price by fifty times.
| Model | Date | Inputs | Weights | Input price per 1M |
|----|----|----|----|----|
| Qwen 3.6 (27B, 35B-A3B) | April 2026 | text, images, video | open, Apache 2.0 | free, your own hardware |
| Qwen 3.7 Max | May 20, 2026 | text | closed | \$1.65–\$2.50 depending on region |
| Qwen 3.7 Plus | May 26, 2026 | text, images, video | closed | \$0.32 |
| Qwen 3.7 Flash | July 15, 2026 | text, images, video | closed | from \$0.03 |
| Qwen 3.8 Max | August 3, 2026 | text, images, video | promised to be open | \$2.00 |
Max itself has several versions, distinguished by the date in the name. This causes separate confusion. `qwen3.7-max` without a date is the May build, and it understands only text. Images and video were added in version `qwen3.7-max-2026-06-08`. Half the reviews call Max multimodal without mentioning that the base version cannot understand images.
If you choose Qwen to save money, the smaller models save more than the top-tier one. Alibaba's $50-a-month subscription confirms the same thing: `qwen3.7-max` is not included at all; only Plus is available.
## Where can you download open Qwen 3.7 weights at 27B and 35B?
Highlights
Nowhere. Alibaba has not released open weights for the 3.7 generation at any size; there is not a single 3.7 model in its official Hugging Face organization. The latest open models are from the previous generation: Qwen3.6-27B from April 21, 2026, and Qwen3.6-35B-A3B from April 15, both under the Apache 2.0 license.
This section is here because many people look for these weights and find nothing. They come looking for model files that don't exist and leave empty-handed.
You can check this with one command, without taking anyone's word for it:
```bash
curl -s "https://huggingface.co/api/models?author=Qwen&limit=1000" \
| jq -r '.[].id' | grep -iE "3\.7|3\.8"
```
The result is empty: the [Qwen organization on Hugging Face](https://huggingface.co/Qwen) has no models from the 3.7 generation. The only repository with a similar name belongs to a third-party user and is not affiliated with Alibaba.
So the choice today is:
1. **You need model files on your own machine** — get Qwen3.6-27B or Qwen3.6-35B-A3B under the Apache 2.0 license, with no payments. Their context window is 262,144 tokens and can stretch to nearly a million.
2. **You need the top-tier 3.7 Max model** — paid API access only; there is no way to run it locally.
3. **You can wait a week** — Alibaba has promised to open the weights for Qwen 3.8 Max and a separate 27-billion-parameter model. More on this below in the section on 3.8.
The difference between 27B and 35B-A3B, if the names aren't clear. In the first, all 27 billion parameters work on every response: in compressed form, that's about 17 GB, which fits on one gaming graphics card. The second is built differently: it has 35 billion parameters, but about three billion are activated per response, so it's faster at the same quality.
## What can Qwen 3.7 Max do according to the official specifications?
Highlights
The context window is one million tokens, roughly an entire project plus its documentation. The model can return an answer up to forty pages long and can think through just as much in advance. It can call tools, search the web, make repeat submissions of text cheaper, and remember its reasoning between steps. It cannot do two things: respond strictly according to a specified schema or adapt to your data.
The figures below come from Alibaba Cloud documentation.
| Parameter | Value |
|---------------------------------|-------------------|
| Context window | 1,000,000 tokens |
| Maximum input | 991,808 tokens |
| Maximum input with reasoning | 983,616 tokens |
| Maximum response | 65,536 tokens |
| Maximum reasoning | 262,144 tokens |
| Tool calls | yes |
| Web search | yes |
| Context caching | yes |
| Strict response format | no |
| Adaptation to your data | no |
A million tokens is roughly 700,000 words—that is, an entire project with its documentation. It sounds like a solution to all memory problems, but in practice quality starts to decline well before that mark; see the section on where the model falls short.
About the two “no” entries in the table. A strict response format means the model must return data exactly according to a specified schema, without improvisation. Without it, any automation that passes the answer on to a program needs a safeguard: you'll have to check the format yourself. You also can't adapt the model to your own data, so working through context remains the option.
Alibaba does not disclose the model's parameter count anywhere. For the released 3.8 Max, it has at least named the total size—2.4 trillion—but it has not disclosed the active portion either.
## How much does Qwen 3.7 Max really cost?
Highlights
One model is sold at different prices, depending on the region. In Alibaba's Singapore region, it costs $2.50 per million input tokens and $7.50 per million output tokens; in Beijing, $1.65 and $4.95. Alibaba's prices are also subject to a temporary 50% discount. A free million tokens are available only in the Singapore region for 90 days.
Reviews get the prices confused because they take the figure for one region and present it as universal. Here are both as of August 3, 2026. The third row shows the generation's mid-tier model for comparison:
| Where | Input per 1M | Output per 1M | Cache |
|----|----|----|----|
| Alibaba Cloud, Singapore | \$2.50 (\$1.25 with discount) | \$7.50 (\$3.75 with discount) | \- |
| Alibaba Cloud, Beijing | \$1.65 (\$0.83 with discount) | \$4.95 (\$2.48 with discount) | \$0.33 for repeated text |
| Qwen 3.7 Plus, Singapore | \$0.40 | \$1.60 | \- |
The discount is marked temporary, but Alibaba does not state an end date. Batch mode in the Beijing region cuts the price in half, but the response does not arrive right away.
There are three catches with free access, and all three undermine what you may read in other reviews:
1. **One million free tokens** are granted per model for 90 days from activation, and only on Alibaba's international platform, which operates through Singapore. There is no free quota in Beijing.
2. **Free Qwen Code access through a Qwen account was disabled on April 15, 2026.** The wording in the documentation is unequivocal. Any material promising free access through a Qwen account describes something that no longer exists.
3. **Two hundred free requests per day** to 3.7 Max were offered through a separate Alibaba product called Qoder. That promotion ended on July 31, 2026, three days ago.
The token price and the cost of a finished task are different quantities. Measure input and output tokens across all attempts on the same task; verbose output and retries can outweigh a low listed rate. The 100-million-versus-63-million output comparison lacks an attributable primary measurement, so it is not a verified benchmark in this article.
In the listed regional price examples, output costs three times as much as input. This makes output volume important to the bill, without establishing that Qwen is more verbose than another model on every task.
A cost-and-accuracy comparison needs a primary platform page, task set, and archived run. Without them, accuracy percentages, dollar totals, and an 87-fold cost claim are not reproducible evidence for a decision. Run the same real task on both candidates and count the cost of all attempts instead.
## Is it true that Qwen 3.7 Max beats Claude?
Highlights
In its May announcement, Alibaba showed it ahead of Claude Opus 4.6 on 14 of 22 benchmarks. In its August blog post about Qwen 3.8, the same company remeasured 3.7 Max on updated benchmark sets, where it trailed Opus 4.8 by ten points or more. An independent rank needs an attributable primary result before it can be treated as confirmed. Alibaba’s own comparisons must remain labelled as vendor-reported figures.
A human-voted ranking needs the platform name, exact model variant, score, vote count, and snapshot date. Without a primary link or saved ranking snapshot, the figures do not establish Qwen’s position against Claude, Kimi, or any other model.
A separate caveat concerns Alibaba's May announcement, which compared the model with Claude Opus 4.6. Those figures still circulate in reviews, but they can't be checked today: the announcement page on the Qwen site returns an empty shell with no figures at all, and the mirror in the Alibaba Cloud community returns the same. Those numbers are not included in the table below.
Before using an independent comparison to choose a model, check its primary result and measurement setup. The reference dates below are unverified and should not be treated as confirmed measurement dates:
| Comparison to verify | Unverified reference date | Evidence missing here | Decision before verification |
|----|----|----|----|
| Intelligence index | August 3, 2026 | Primary platform and snapshot not supplied | No score or rank claimed |
| Output-speed comparison | August 3, 2026 | Primary measurement and serving setup not supplied | Measure on the intended provider |
| Cost-and-accuracy index | July 31, 2026 | Platform, task set, and run receipt not supplied | No comparative cost or accuracy claimed |
| Terminal agent leaderboard | August 3, 2026 | Exact leaderboard snapshot not supplied | No absence or rank inferred |
| Human-voted coding ranking | August 1, 2026 | Platform/model variant and snapshot not supplied | No score or rank claimed |
Do not fill these gaps with a similarly named model or today’s changing leaderboard. A reproducible independent comparison needs a direct primary measurement and its date; no primary result linked here supports such a numerical comparison.
In its August 3.8 announcement, Alibaba remeasured the May 3.7 Max on fresh benchmark sets: it scored 21.6 versus 56.6 for the new version on DeepSWE 1.1, and 40.7 versus 73.5 on FrontierSWE. That's a 2.5-fold gap in eleven weeks, shown not by an independent platform but by the manufacturer itself.
Response speed is worth measuring on the provider and model snapshot you will actually use. Record time to first output, sustained output speed, and whether the task finished correctly. The original unattributed 206-token-per-second comparison does not establish the latency of your setup.
If Claude is unavailable, Qwen can be a candidate for a bounded pilot. The vendor table and unattributed historical notes do not establish an equal replacement or a universal performance gap.
## Where does Qwen 3.7 fall short in real work?
Highlights
The same four complaints recur from different people: quality drops well before the promised million tokens, Alibaba's cloud times out under load, the program loses access to files without saying so, and Qwen understands vague instructions less well than Claude. The last issue hits non-programmers hardest.
What follows are complaints from real people. There are no benchmarks here, and the claims can't be verified, so they are presented as complaints. Some are about Alibaba's cloud and tools generally and have nothing to do with a specific model version.
**A million tokens are advertised, but not all of them work well.** People said this directly about the previous generation of Qwen: the closer you get to a million, the slower the responses. No verified benchmarks of long-context quality for 3.7 Max were found, so for now this is only an expectation.
**The cloud can't handle the load.** On the day 3.8 came out, August 3, user discussions filled with similar complaints.
The same discussion describes access being denied despite a paid plan: with both a subscription and a prepaid-balance key, the tool returns an authorization error.
**The program reports success when it hasn't succeeded.** The same day, users reported that the Qwen desktop app said everything worked and gave a link to a folder that did not exist.
Without experience, you won't catch that: the agent says “done,” but the disk is empty.
**The models need detailed instructions.** Qwen understands vague instructions less well than Claude. This divides users by type: a programmer will write a technical specification and get a good result for little money; an entrepreneur will say “make it look nice” and get something else.
There are positive reviews too; without them, the picture would be one-sided. Someone who worked with a preview version of 3.8 for two weeks says it completed as many tasks in their private repository as Fable 5, and they don't miss the top-tier models. Qwen is praised on small projects and criticized on large ones.
## How can you try Qwen 3.7 without installing it?
Highlights
There are three options. The chat on the Qwen website runs in a browser and requires nothing. API access through Alibaba Cloud Model Studio connects to any program, including Claude Code. Alibaba's own terminal tool is called Qwen Code; it installs with one command and can work with models from different providers.
It's best to start in the browser. Open the chat, give the model a real task of yours, and look at the response. Five minutes can save you an evening of setup if the model isn't a good fit.
Here are the options in order of increasing complexity:
1. **Browser chat.** Nothing to install; registration is straightforward. Good for getting a feel for its responses.
2. **API access.** Create a key in Alibaba Cloud Model Studio, then add it to your program. This lets you use the model in an editor, a bot, or your own automation.
3. **Qwen Code in the terminal.** Alibaba's open-source tool, under the Apache 2.0 license, installs with `npm install -g @qwen-code/qwen-code@latest` if Node.js 22 or later is installed. It can use not only Qwen but also models from OpenAI, Anthropic, and Google, as well as local models through Ollama and vLLM.
There is another option: Alibaba has an API compatible with Anthropic, and Claude Code can connect directly to Qwen. The instructions are in the Alibaba Cloud documentation:
```bash
npm install -g @anthropic-ai/claude-code
export ANTHROPIC_BASE_URL=https://coding-intl.dashscope.aliyuncs.com/apps/anthropic
export ANTHROPIC_AUTH_TOKEN=
export ANTHROPIC_MODEL="qwen3.7-plus"
claude
```
The model on the last line is Plus, not a typo: Alibaba's subscription does not include the top-tier Max at all; it is billed separately based on usage.
The familiar interface stays the same, but the model inside changes. It's hard to find a cheaper way to compare: one task, one program, one evening.
One caveat that other instructions usually forget: the key on the last line is from Alibaba; you can't put an Anthropic key there. And with this configuration, your requests and project snippets go to Alibaba's servers, not Anthropic's. What happens to them there is covered in the section on data below.
Step-by-step installation of Qwen Code, including setting up a key, choosing a plan, and troubleshooting error 401, is covered separately in [the article about Qwen Code as a Claude Code replacement](https://ailoot.ai/blog/en/qwen-code-token-plan-setup/). It is not repeated here.
## Access and payment terms
Highlights
Alibaba Cloud operates in six regions, including Frankfurt and Singapore; its pages do not provide an official list of countries with restrictions. Access requires an Alibaba Cloud account and a payment method accepted by that account.
Payments deserve a separate article, and there will be no advice here on circumventing anything.
Here is what can be said as fact:
- Regions: Beijing, Hong Kong, Singapore, Tokyo, Frankfurt, Virginia. The international version of the service operates through a Singapore address.
- An Alibaba Cloud account and an access key from the Model Studio console are required.
- It is worth checking your bank's terms before paying.
A separate note about materials promising a ready-made list of payment methods: they usually lead to unvetted intermediaries and end with money being sent nowhere. That's why there are no such links here.
To compare alternatives, see [the GLM-5.2 review](https://ailoot.ai/blog/en/glm-5-2-vs-claude-code/) and the comparison of DeepSeek and Claude Code. Each has its own access story and price.
## What happens to your data in Alibaba Cloud?
Highlights
Alibaba Cloud says it does not use customer data to train models and encrypts data in transit to the AES-256 standard. A separate document dated March 15, 2026, uses softer wording: no use without explicit consent. The retention period for requests and responses is not published; only the period for service logs is specified—30 days.
The question “Will my data leak?” is addressed below using official documents; reviews are not suitable for this.
The first document states directly:
Alibaba Cloud protects data privacy and never uses your data to train models.
— Alibaba Cloud Model Studio, https://www.alibabacloud.com/help/en/model-studio/what-is-model-studio
The second document, about data used for training, says the same thing differently: without the customer's explicit consent, data is not used to develop models. The difference between “never” and “without explicit consent” is significant, so both statements are quoted as written.
| Question | Answer according to the documents |
|----|----|
| Training on customer data | stated not to occur |
| Data-in-transit encryption | AES-256 |
| Certification | SOC 2 with an unqualified opinion |
| Where data is stored | in the region selected when the workspace is created |
| Request retention period | not published |
| Service-log retention period | 30 days by default |
The region is selected once and determines the data-storage jurisdiction: Frankfurt and Singapore are not the same as Beijing. But the retention period for the requests themselves is not in the public documents. That's a gap, and it shouldn't be ignored.
Don't send anything you can't show to outsiders to someone else's model, whatever the provider. That includes access keys, customer data, and contracts.
## What changes with Qwen 3.8, released on August 3?
Highlights
Qwen 3.8 Max was released on August 3, 2026. The original 4th-versus-23rd human-voted ranking comparison has no attributable primary snapshot in the supplied sources and is not used here. It has 2.4 trillion parameters; Alibaba does not say how many are active for each response. Alibaba promises to open the weights next week, making it the first open top-tier model in its lineup.
The model you're reading about stopped being the flagship yesterday.
| Feature | Qwen 3.7 Max | Qwen 3.8 Max |
|----|----|----|
| Date | May 20, 2026 | August 3, 2026 |
| Parameters | undisclosed | 2.4 trillion, active portion undisclosed |
| Inputs | text; images and video in the June 8 version | text, images, video |
| Maximum response | 65,536 tokens | 131,072 tokens |
| Strict response format | no | yes |
| Independently verified coding rank in supplied sources | not established | not established |
| Open weights | no | promised for next week |
Alibaba announced the open weights separately: this is the first time it has opened the weights of a top-tier model, and they are expected next week. Along with 3.8 Max itself, a separate 27-billion-parameter model is promised—one you can keep on your own hardware.
Alibaba claims 86.6 on Terminal-Bench 2.1, compared with 84.6 for Claude Opus 4.8. There are two reasons to pause. Some comparisons in the announcement are against Qwen3.7-Plus, a mid-tier model from the previous generation, instead of its own top-tier Max, which makes the improvement look larger than it is. This article’s supplied sources do not establish an independently verified 3.8 benchmark or vote count. That is a limit of the evidence here, not proof that no independent evaluation exists.
What does this mean in practice?
1. **Considering Qwen as a work model?** Look at 3.8 right away. There is little reason to pay for 3.7 today: in the Singapore region, 3.8 is even cheaper, at $2 versus $2.50 for input.
2. **Need open weights?** Wait a week. Both 3.8 Max itself and a separate 27-billion-parameter model for your own hardware have been promised.
3. **Already working on 3.7 and happy with it?** Don't rush to switch. This article does not establish a verified independent head-to-head comparison. Test a replacement on a representative task before changing a working setup.
The story repeats about once a quarter: [Kimi K3](https://ailoot.ai/blog/en/kimi-k3-vs-claude-code/), [GLM-5.2](https://ailoot.ai/blog/en/glm-5-2-vs-claude-code/), and DeepSeek came with the same promises. Chasing every release costs more than setting things up once so you can change models with one line in the settings.
## Who is Qwen 3.7 right for, and who is it not for?
Highlights
For a pilot, choose a clearly specified task and measure response speed, completion, and rework. Broad autonomous work needs its own acceptance check; neither an unattributed ranking nor the price per token establishes which model will perform better on your project.
Here's a brief breakdown by work type to help you decide.
| Task | Is Qwen worth using? |
|----|----|
| Repetitive edits with a clear description | useful first pilot; measure result and cost |
| Large volumes of simple requests | pilot a small sample before scaling |
| Analyzing a large project as a whole | verify retrieval and correctness on the required context |
| Long stretches of independent agent work | head-to-head performance is not established here |
| A task that needs interpretation | test the same acceptance criteria on both candidates |
| Work with sensitive data | no; request retention period is unpublished |
The evidence here supports a pilot rather than a definitive everyday-work ranking against Claude Code. Try clearly specified repetitive tasks first and inspect whether the output and total cost meet your needs. That's exactly how to use [free and cheap models](https://ailoot.ai/blog/en/free-coding-models/): keep the heavy work on the top-tier model and send the stream of routine tasks to the cheaper one.
Before switching models, calculate how much you pay for rework. If the agent rewrites the same thing three times, the rework costs more than the tokens. Switching models won't fix that; you need to look at what's in its context.
### Sources
- [Qwen3.7-Max specifications in Alibaba Cloud Model Studio](https://www.alibabacloud.com/help/en/model-studio/qwen3-7-max) — context window, limits, versions, and snapshots
- [Alibaba Cloud Model Studio model pricing](https://www.alibabacloud.com/help/en/model-studio/model-pricing) — regional prices, discount, caching
- [What is Alibaba Cloud Model Studio](https://www.alibabacloud.com/help/en/model-studio/what-is-model-studio) — regions, access addresses, statement about training on data
- [Model Studio Privacy Notice](https://help.aliyun.com/en/model-studio/privacy-notice) — encryption, certification, data processing
- [Qwen and Wan training data disclosure](https://www.alibabacloud.com/help/en/model-studio/qwen-and-wan-training-data-disclosure) — wording about explicit consent
- [Connecting Claude Code to Alibaba models](https://www.alibabacloud.com/help/en/model-studio/claude-code) — working address and environment variables
- [Qwen3.8-Max specifications](https://www.alibabacloud.com/help/en/model-studio/qwen3-8-max) — parameters, prices, strict response format
- [Qwen organization on Hugging Face](https://huggingface.co/Qwen) — which models have actually been released openly
- [Qwen3.6-27B model card](https://huggingface.co/Qwen/Qwen3.6-27B) — Apache 2.0 license, size, context
- [Qwen Code repository](https://github.com/QwenLM/qwen-code) — installation, license, supported providers
- [Qwen Code authentication documentation](https://qwenlm.github.io/qwen-code-docs/en/users/configuration/auth/) — free-tier cancellation on April 15, 2026
More materials are available on [Ailoot](https://ailoot.ai/).
## Questions about Qwen 3.7 and 3.7 Max: what the model for coding can do in 2026
### Do the cited independent ranks and cost comparisons establish Qwen’s position?
No. No primary measurement links or archived snapshots are linked here for those figures. Alibaba’s reported figures remain vendor claims, and an independent comparison needs an attributable source and date before it can establish a score, rank, or cost advantage.
### How can I check whether Qwen is cheaper for my actual coding task?
Use the same task and acceptance checks for each candidate. Record input and output usage, retries, time to a working result, and the provider’s regional rates; compare total cost per accepted result rather than the listed input-token price alone.
## Ailoot · The company and its project portfolio
Ailoot combines established operating experience, a large client base, stable service and ongoing KPI control.
The company and its project portfolio: millions of customers; monthly sales in the billions.
Ailoot brings that operating experience to API access for AI features in customer products, agents, applications and automations.