Last verified: August 13, 2026. Qwen AI cloud hosting is not one product or deployment method. You can call the production qwen3.8-max service through a supported QwenCloud or Alibaba Cloud route, use a third-party managed provider, or operate an exact official open-weight checkpoint in your own cloud account. Hosted service IDs and downloadable repository IDs are separate artifacts with different pricing, licensing, lifecycle, and operational responsibilities. Review the Qwen3.8 Max guide before selecting the production route.
Independent-guide notice: qwen-ai.chat is an independent informational website and is not Qwen, Alibaba Cloud, or Alibaba Group. This page is not endorsed by those organizations. We verified model names, endpoint formats, access conditions, and licenses against first-party documentation and model repositories available on the date above. Provider catalogs, aliases, prices, quotas, and preview access can change; confirm them before production use. We did not substitute vendor benchmark claims for testing on your workload.
Quick answer: which Qwen hosting route should you choose?
| Route | What you deploy | Best fit | Main trade-off |
|---|---|---|---|
| Alibaba Cloud Model Studio | A hosted API model identified by a Model Studio model ID | Current first-party hosted Qwen models, official regional endpoints, and minimal infrastructure work | Availability, aliases, pricing, and features vary by region and API |
| Third-party managed inference | A provider endpoint for a Qwen checkpoint or provider-specific model alias | Existing cloud commitments, a multi-model platform, or managed dedicated capacity | Model IDs, quantization, features, regions, retention, and lifecycle are provider-specific |
| Self-hosting | Files from an official open-weight repository, served by your chosen runtime | Maximum infrastructure control, custom networking, or reproducible model revisions | You own sizing, security, upgrades, scaling, reliability, and evaluation |
Do not treat these routes as interchangeable. A hosted API name such as qwen3.7-plus is not a downloadable checkpoint ID. A repository ID such as Qwen/Qwen3.6-35B-A3B identifies open weights, but it does not guarantee that every managed provider serves the same revision, precision, context limit, or tool behavior.
Current Qwen model identity: hosted IDs versus open checkpoints
| Name or ID | Verified status at cutoff | Deployment meaning |
|---|---|---|
qwen3.8-max |
Production flagship released August 3, 2026 | Hosted service. QwenCloud publishes 1M context, text/image/video input, text output, hybrid thinking, and $2 input / $6 output per 1M tokens for its Marketplace route. No public weights were verified. |
qwen3.8-max-preview |
Retired Token Plan input string | Old calls are temporarily accepted but automatically routed to production qwen3.8-max. Update configurations and do not treat it as a separate preview runtime. |
qwen3.7-max |
Older hosted Max entry still listed where supported | No longer the current flagship; no reviewed notice assigns it a retirement date. |
qwen3.7-plus and qwen3.7-flash |
Balanced and lower-cost hosted starting points | Compare measured quality, route, region, tools, and price. |
Qwen/Qwen3.8-2.4T-A95B and its official FP8 variant |
Official open-weight Qwen3.8 checkpoints under the custom Qwen3.8-Max License | Self-hostable with substantial infrastructure; text-only and thinking-only, and not aliases for hosted Qwen3.8 Max. |
Qwen/Qwen3.6-35B-A3B and Qwen/Qwen3.6-27B |
Smaller official open-weight checkpoints | Self-hostable under their repository licenses; not aliases for Qwen3.8 Max. |
QwenCloud exposes production qwen3.8-max. The retired qwen3.8-max-preview string is temporarily accepted on the documented Token Plan route but automatically routes to production. Separately, Qwen publishes the open checkpoint Qwen/Qwen3.8-2.4T-A95B. Record the exact product, region, API protocol, base URL, key type, repository or service ID, and returned model ID; the hosted service and open checkpoint have different feature surfaces.
Route 1: QwenCloud or Alibaba Cloud Model Studio
The current QwenCloud and Alibaba Cloud model selectors recommend qwen3.8-max for the strongest reasoning tier. QwenCloud’s Responses reference explicitly lists the production ID; its Marketplace also publishes a DashScope multimodal example. Availability, base URLs, deployment scopes, pricing, and credentials still depend on the selected QwenCloud product or Alibaba Cloud workspace.
This site did not complete an authenticated Qwen3.8 request during the August 3 update. Keep the existing Qwen3.7 Plus sample as a documentation-based route example until a small production request succeeds. Record the exact region, route, protocol, model ID, status, request ID, usage, settings, and output without exposing the key.
Workspace endpoints for the OpenAI-compatible Responses API
| Region | Verified base URL |
|---|---|
| Singapore | https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1 |
| China (Beijing) | https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/compatible-mode/v1 |
| US (Virginia) | https://dashscope-us.aliyuncs.com/compatible-mode/v1 |
| China (Hong Kong) | https://{WorkspaceId}.cn-hongkong.maas.aliyuncs.com/compatible-mode/v1 |
| Germany (Frankfurt) | https://{WorkspaceId}.eu-central-1.maas.aliyuncs.com/compatible-mode/v1 |
| Japan (Tokyo) | https://{WorkspaceId}.ap-northeast-1.maas.aliyuncs.com/compatible-mode/v1 |
Alibaba recommends workspace-specific domains where documented. It also warns that the older Responses path ending in /api/v2/apps/protocols/compatible-mode/v1/responses is being deprecated; new integrations should use the base URLs above and the /responses resource. Use the API key created for the correct region and workspace.
Node.js example
This example uses the Singapore workspace endpoint and a currently documented hosted ID. Keep the key on the server, pin the OpenAI SDK version in package.json, and commit the lockfile.
import OpenAI from "openai";
const workspaceId = process.env.DASHSCOPE_WORKSPACE_ID;
const apiKey = process.env.DASHSCOPE_API_KEY;
if (!workspaceId || !apiKey) {
throw new Error("Missing Model Studio workspace ID or API key");
}
const client = new OpenAI({
apiKey,
baseURL:
`https://${workspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1`,
});
const response = await client.responses.create({
model: "qwen3.7-plus",
input: "Summarize the deployment risks in five bullet points.",
});
console.log(response.output_text);Region is not the same as data residency
Separate four decisions: the hostname receiving the request, the workspace and API-key region, the model’s documented deployment scope, and the contractual rules for processing, storage, and logs. Alibaba’s model tables use scopes such as Chinese mainland, International, Global, EU, US, and Hong Kong, and the scope can differ by model within one regional table. Do not infer legal residency from the endpoint name alone. Confirm the exact model row, service terms, retention controls, and any required data-processing agreement.
Token Plan is interactive tooling, not application hosting
Token Plan Individual and Team support production qwen3.8-max; calls using the retired qwen3.8-max-preview string are routed to it. The subscription is limited to interactive use in compatible programming and agent tools. QwenCloud prohibits automated scripts, application backends, and non-interactive batch processing. A compatible request shape does not override those terms.
Use QwenCloud pay as you go for an application or API integration. Keep Token Plan, PAYG, Coding Plan, and Model Studio credentials and base URLs separate.
Route 2: third-party managed inference
A managed inference provider can operate an official open checkpoint for you or expose a provider-specific Qwen alias. Examples of primary catalogs include Together AI, Fireworks AI, and Amazon Bedrock. Catalog presence is not proof of equivalence to Alibaba Cloud’s hosted model. Before selection, verify the exact provider model ID, upstream checkpoint and revision, precision or quantization, context and output limits, reasoning and tool support, dedicated versus shared capacity, rate limits, regions, retention, deprecation policy, and commercial terms.
Run the same evaluation set across providers. A familiar OpenAI-compatible request format only standardizes part of the interface; tokenization, optional parameters, error behavior, streaming events, safety layers, and model implementation may still differ. Keep a provider adapter in your application instead of scattering model IDs and base URLs through business logic.
Route 3: self-host official open-weight Qwen checkpoints
For the highest-capability current open Qwen release, start from Qwen/Qwen3.8-2.4T-A95B or its official FP8 variant. This 2.4T-total, 95B-active checkpoint is text-only, requires thinking mode, and uses the custom Qwen3.8-Max License; plan storage, accelerator memory, parallelism, and startup time from measured requirements for the full checkpoint. For smaller multimodal self-hosting, consider Qwen/Qwen3.6-35B-A3B or Qwen/Qwen3.6-27B under their repository licenses. Read each exact model card, license file, configuration, and serving-framework compatibility before deployment.
Qwen documents Transformers serving, SGLang, and vLLM as possible runtimes. That does not create one universal GPU count or throughput number. Memory and performance depend on the exact checkpoint, weight and cache precision, context length, batch policy, concurrent sequences, prompt distribution, output length, runtime build, kernels, accelerator generation, interconnect, and service-level target. Profile your intended workload and keep headroom for failures and traffic bursts.
Pin every production layer
runtime_image: registry.example/inference-runtime:<tested-version>@sha256:<digest>
model_id: Qwen/Qwen3.6-35B-A3B
model_revision: <repository-commit>
config_revision: <internal-config-commit>
The values above are placeholders, not vendor tags. Select a runtime release that officially supports the chosen checkpoint, test it, and pin the exact image digest. Pin the model repository revision and verify downloaded file hashes. Never deploy an unversioned :latest image. Record drivers, libraries, tokenizer, chat template, reasoning parser, and quantization artifacts so a rollback can reproduce the previous service.
Security and production controls
- Keep provider keys in a secret manager; use separate credentials per environment, least privilege, rotation, and rapid revocation.
- Keep model calls server-side. Apply authentication, tenant isolation, quotas, request-size limits, timeouts, retry budgets, and egress controls at your gateway.
- Classify prompts and attachments before sending them. Minimize sensitive data and verify provider retention, training-use, deletion, encryption, and audit terms.
- Allowlist tools and validate tool arguments. Treat model output as untrusted input before database queries, shell actions, file access, or outbound requests.
- For self-hosting, scan images and dependencies, maintain an SBOM, patch drivers and runtimes, restrict model-download sources, and isolate inference nodes.
- Use redaction and sampling for logs. Do not store API keys, complete private prompts, reasoning traces, or personal data by default.
When managed-provider cost is part of the hosting decision, compare like-for-like routes in the AI cloud pricing comparison before estimating production spend.
Observability, cost, and capacity planning
For every request, capture a correlation ID, provider, region, exact model ID, alias or snapshot, deployment revision or image digest, prompt-template version, input/cached/output token counts, time to first token, end-to-end latency, status, retries, rate-limit events, tool calls, safety outcome, and estimated cost. Use hashed or redacted payload references when content logging is not justified.
Do not publish a timeless GPU price or tokens-per-second claim. For hosted APIs, estimate monthly cost as: input tokens × input rate + cached tokens × cache rate + output tokens × output rate + tool, batch, or provisioned-capacity charges. Apply the provider’s current tiers and region. For self-hosting, include accelerators, CPU and RAM, storage, network egress, orchestration, observability, redundancy, idle capacity, engineering, security, and on-call work. Compare total cost at measured utilization and required availability.
Evaluation checklist before launch
- Freeze a representative, permission-safe test set with expected outcomes and failure cases.
- Score task quality, groundedness, instruction adherence, structured output, tool-call validity, refusal behavior, multilingual needs, and safety.
- Measure latency distributions, time to first token, throughput, errors, rate limits, recovery, and cost under realistic concurrency and prompt lengths.
- Test prompt injection, data leakage, cross-tenant isolation, unsafe tool arguments, oversized inputs, malformed documents, and network failures.
- Record the exact provider, region, API, model ID or repository revision, runtime, and configuration for each result.
- Canary the winning configuration, define rollback thresholds, and retest after any alias, runtime, prompt, safety-policy, or infrastructure change.
Model migration and lifecycle
Hosted aliases can be repointed: Alibaba’s pricing documentation, for example, records the dated snapshot currently associated with several aliases. Use a dated model ID when reproducibility matters and the region supports it. Keep an alias-to-snapshot registry, a tested fallback, deprecation alerts, and a rollback window. During migration, compare answer quality, reasoning settings, JSON schemas, tool calls, token use, latency, safety behavior, and errors—not only whether the request returns HTTP 200.
Frequently asked questions
Is Qwen AI cloud hosting the same as Qwen Chat?
No. Qwen’s consumer web experience is an application. Cloud deployment concerns APIs or open checkpoints integrated into your own system.
Can I self-host Qwen3.8 Max or Qwen3.7 Max?
Qwen now publishes the related Qwen/Qwen3.8-2.4T-A95B checkpoint and an official FP8 variant. These are not downloadable copies of every hosted qwen3.8-max feature: the open checkpoint is text-only and thinking-only, while Qwen says the hosted service adds vision input, non-thinking mode, a 1M default context, and official built-in tools. qwen3.7-max remains a hosted service ID without a matching checkpoint established here. Review the custom Qwen3.8-Max License and size infrastructure for the full 2.4T-parameter checkpoint.
Which current Qwen models can I self-host?
Verified current examples are Qwen/Qwen3.6-35B-A3B and Qwen/Qwen3.6-27B, both released through official repositories under Apache License 2.0. Earlier official Qwen3 checkpoints remain options where their model cards and licenses fit your use case.
How many GPUs does Qwen3.6 require?
There is no responsible universal answer. Determine it from the exact checkpoint, precision, context, concurrency, runtime, latency target, and redundancy plan, then validate with a workload benchmark.
Can I use the OpenAI Node.js SDK?
Yes, Model Studio documents an OpenAI-compatible Responses API. Configure the correct regional base URL, a server-side DashScope key, and a supported model ID. Compatibility does not mean every OpenAI-specific feature behaves identically.
Does choosing Frankfurt guarantee EU-only data processing?
Do not assume so from the endpoint. Check the exact model’s deployment scope and the current service, logging, retention, subprocessors, and contractual terms for your account.
Is third-party Qwen inference identical to Model Studio?
No. A provider may serve an open checkpoint, a quantized variant, or its own alias with different limits and operational controls. Verify and benchmark the exact offering.
Continue planning your Qwen deployment
- Use the Qwen API guide to verify endpoints, model IDs, request formats, streaming, tools, and provider differences.
- Compare regional and provider costs in the Qwen API pricing guide.
- Review the Qwen models hub before selecting a hosted alias or open-weight checkpoint.
- Use the Qwen download guide to separate official apps, model weights, and local runtime requirements.
Official and primary sources
- Alibaba Cloud Model Studio: current text-generation model matrix
- Alibaba Cloud Model Studio: OpenAI-compatible Responses API, models, and regional endpoints
- Alibaba Cloud Model Studio model availability, deployment scopes, alias notes, and current pricing
- Official Qwen3.6 repository, checkpoint IDs, deployment frameworks, and license statement
- Official Qwen3.6-35B-A3B model card and files
- Official Qwen3.6-27B model card and files
- Official Qwen3 model collection
- QwenCloud text-generation model catalog
- QwenCloud Token Plan and Qwen3.8 preview conditions
- Together AI serverless model catalog, Fireworks AI model guidance, and Amazon Bedrock Qwen model cards for provider-specific verification