Last verified: August 13, 2026.
Independent and unofficial: Qwen-AI.chat is not operated by, affiliated with, or endorsed by Alibaba Cloud, QwenCloud, Alibaba Group, or the Qwen team. The specifications and prices below were checked against first-party pages on the date above. This site did not complete an authenticated qwen3.8-max inference request during this update, so no API example or model result on this page is described as independently live tested.
Qwen3.8 Max is QwenCloud’s current production flagship under the exact model ID qwen3.8-max. QwenCloud announced it on August 3, 2026 as a native vision-language Mixture-of-Experts model with 2.4 trillion total parameters, hybrid thinking enabled by default, and a 1M-token context category.
QwenCloud’s current Token Plan documentation states that qwen3.8-max-preview has ended its preview period and is officially retired. The old model ID remains callable for compatibility, but requests are automatically routed to the production qwen3.8-max model; Credits deduction and usage statistics are also calculated as qwen3.8-max. Update configurations to the production ID rather than relying on the retired string.
Qwen has also published the official open-weight checkpoints Qwen/Qwen3.8-2.4T-A95B and Qwen/Qwen3.8-2.4T-A95B-FP8. They are distinct from the hosted qwen3.8-max service. The open checkpoint is text-only, requires thinking mode, has a native context length of 262,144 tokens that the model card says can be extended to 1,010,000, and is governed by the custom Qwen3.8-Max License rather than Apache 2.0. Qwen’s model card states that the hosted Qwen3.8-Max service is based on this checkpoint but adds vision input, non-thinking support, a 1M context length by default, official built-in tools, and other managed-service features.
Qwen3.8 Max at a glance
| Item | Verified production information |
|---|---|
| Official display name | Qwen3.8 Max |
| Exact production ID | qwen3.8-max |
| Release date | August 3, 2026 |
| Service status | Current production flagship on QwenCloud’s model selector and Marketplace |
| Architecture | Mixture-of-Experts, according to QwenCloud |
| Parameter count | 2.4 trillion total parameters, vendor reported |
| Inputs | Text, images, and video on the QwenCloud Marketplace route |
| Output | Text |
| Context | 1M category; Marketplace publishes 991K maximum input |
| Maximum output | 131K on the current Marketplace route |
| Thinking | Hybrid thinking, enabled by default; Marketplace publishes 262K maximum reasoning |
| Function Calling | Listed as supported by QwenCloud |
| Built-in tools | Listed as supported; exact tools and protocol are route-specific |
| Structured output | Listed as supported for production qwen3.8-max; validate output in application code |
| QwenCloud PAYG list price | $2 input and $6 output per 1M tokens |
| Token Plan | qwen3.8-max is supported. The retired qwen3.8-max-preview string remains callable but is automatically routed to production; Token Plan remains restricted to interactive use in supported tools. |
| Related open weights | Official repositories: Qwen/Qwen3.8-2.4T-A95B and Qwen/Qwen3.8-2.4T-A95B-FP8. These checkpoints are distinct from the hosted service and use the custom Qwen3.8-Max License. |
Limits, rate limits, features, and prices belong to the named QwenCloud route and can change. They are not automatically transferable to another Alibaba region, Qoder, Fireworks AI, or a third-party deployment.
Production qwen3.8-max versus qwen3.8-max-preview
| Exact ID | Lifecycle and access | Safe interpretation |
|---|---|---|
qwen3.8-max |
Production flagship released August 3, 2026; listed by QwenCloud Marketplace, the current model selector, the Responses reference, and both Token Plan allowlists | Use this exact ID when the selected QwenCloud route and account expose the production model. Confirm the key, base URL, protocol, and region together. |
qwen3.8-max-preview |
Retired Token Plan model ID; the preview period has ended | Calls using the old ID remain accepted but are automatically routed to qwen3.8-max, with Credits and usage statistics calculated as production. Update configurations and do not treat the retired string as a separate evaluation route. |
This automatic routing is explicitly documented for qwen3.8-max-preview on the QwenCloud Token Plan route. It must not be generalized to qwen3-max, qwen3.7-max, or any other older alias. Record the configured and returned model IDs, and test the production model before relying on the compatibility route.
Confirmed capabilities and service limits
QwenCloud describes Qwen3.8 Max as a native vision-language model. Its Marketplace accepts text, image, and video input and returns text. This does not mean that the model generates images, video, or audio; those outputs require separate models or platform services.
The current Marketplace publishes the following service limits for qwen3.8-max:
- Context: 1M tokens.
- Maximum input: 991K tokens.
- Maximum input with thinking: 983K tokens.
- Maximum output: 131K tokens.
- Maximum output with thinking: 131K tokens.
- Maximum reasoning: 262K tokens.
These are provider-published ceilings, not a recommendation to fill the window. Leave room for reasoning, tool results, and final output. Long-context accuracy and latency must be tested on the real workload; a maximum accepted length is not proof that every fact inside a one-million-token prompt will be recalled reliably.
QwenCloud’s current selector lists Function Calling, built-in tools, and structured output for production qwen3.8-max. The Marketplace lists Responses built-in tools including web search, web extraction, code interpreter, text-to-image search, and reverse-image search. Tool availability can depend on the protocol and account. A model proposes tool calls; the surrounding application must validate arguments, enforce authorization, run the tool, and validate the returned data.
Qwen3.8 Max pricing
QwenCloud’s model-specific Marketplace displayed the following public list prices on August 3, 2026. Every amount is in US dollars per 1 million tokens:
| Billing item | Public list price |
|---|---|
| Input | $2.00 |
| Output | $6.00 |
| Implicit-cache input | $0.25 |
| Explicit-cache creation | $2.50 |
| Explicit-cache read | $0.17 |
Example: 100,000 uncached input tokens and 20,000 output tokens would cost (0.1 × $2) + (0.02 × $6) = $0.32, before tool fees, taxes, promotions, retries, or other infrastructure costs.
These are QwenCloud Marketplace prices, not a universal Alibaba Cloud Model Studio regional quote. Do not copy them into another provider’s price table without confirming the exact model, region, deployment scope, key, and billing page. See the Qwen API pricing guide for route-separated pricing and cost formulas.
API status and verification boundary
QwenCloud’s current documentation may still accept qwen3.8-max-preview as an input string, but the Token Plan lifecycle notice clarifies that it no longer identifies a separate preview model: calls are routed to qwen3.8-max. New integrations should use qwen3.8-max directly and verify the exact protocol, endpoint, region, and account access.
No authenticated production request was completed for this page update. Before publishing a code block as tested, run one small request with:
- the exact
qwen3.8-maxID; - the QwenCloud PAYG key for the selected workspace;
- the matching base URL and region;
- the intended protocol, such as Responses, Chat Completions, or DashScope multimodal;
- a non-sensitive text prompt first, followed separately by any required image, video, Function Calling, structured-output, or built-in-tool test.
Record the date, endpoint family, region, HTTP status, request ID, returned model ID, usage fields, settings, and final output without exposing the API key. Until that succeeds, describe code as documentation-based, not independently executed. Continue with the Qwen API guide for route and credential boundaries.
Token Plan is not a backend API plan
Token Plan Individual and Team support qwen3.8-max. The retired qwen3.8-max-preview ID remains callable only through the documented automatic routing to production. The plan still has a decisive use restriction: QwenCloud limits Token Plan to interactive use inside compatible programming and agent tools. Automated scripts, custom application backends, and non-interactive batch processing are prohibited and may lead to suspension or API-key revocation.
Use QwenCloud pay as you go for an application or API integration. Token Plan uses a dedicated key and Token Plan base URL; its credentials must not be mixed with QwenCloud PAYG, Coding Plan, or Alibaba Cloud Model Studio endpoints. Credits and temporary promotions are not a dollar-per-million-token PAYG rate.
Can Qwen3.8 be downloaded or self-hosted?
Yes. Qwen has published Qwen/Qwen3.8-2.4T-A95B and an official FP8 variant. The repositories contain post-trained weights and configuration files for supported inference frameworks such as vLLM, SGLang, and TokenSpeed.
The downloadable checkpoint is not a local copy of every hosted qwen3.8-max feature. It is text-only and thinking-only, while Qwen says the managed Qwen3.8-Max service adds vision input, non-thinking support, a 1M context length by default, and official built-in tools. Read the custom Qwen3.8-Max License before deployment and do not transfer hosted pricing, tools, modalities, limits, or service terms to the self-hosted checkpoint. See the Qwen download and local setup guide and the Qwen cloud hosting guide for the hosted-versus-open distinction.
Qwen3.8 Max versus Qwen3.7 Max
QwenCloud now recommends qwen3.8-max for its highest-capability tier. qwen3.7-max is an older hosted Max entry that remains listed where supported. No reviewed lifecycle notice assigns Qwen3.7 Max a retirement date, so it must not be called retired.
Evaluate both only on the same route and workload. Record model IDs, reasoning settings, input modalities, context lengths, tools, latency, token usage, errors, and success criteria. Vendor statements about improved capability over Qwen3.7 are not a substitute for independent results.
Frequently asked questions
What is the current Qwen3.8 Max model ID?
The production hosted ID is qwen3.8-max. qwen3.8-max-preview is officially retired; the old string remains callable on the documented Token Plan route but is automatically routed to production. Update configurations to qwen3.8-max.
Is Qwen3.8 Max fully released?
Yes, in two distinct distribution lanes. qwen3.8-max is the production hosted QwenCloud service, while Qwen/Qwen3.8-2.4T-A95B and its FP8 variant are official downloadable checkpoints. The hosted service and open checkpoint are related but do not expose identical modalities, thinking controls, tools, or default context behavior.
Is Qwen3.8 Max the same as Qwen3-8B?
No. “3.8” is the generation/version label in Qwen3.8 Max. Qwen3-8B is a separate, earlier checkpoint whose “8B” denotes its approximate parameter scale.
Does Qwen3.8 Max support images and video?
QwenCloud’s Marketplace lists text, image, and video input with text output for the production route. It does not list image, video, or audio generation as outputs.
What does Qwen3.8 Max cost?
The QwenCloud Marketplace list price checked on August 3, 2026 was $2 per 1M input tokens and $6 per 1M output tokens, with separate cache rates. Confirm the live page and account console before budgeting.
Can Token Plan power my application’s backend?
No. Current Token Plan terms limit it to interactive use in compatible programming and agent tools. QwenCloud directs application and API integrations to pay as you go.
Is Qwen3.8 Max open source?
Qwen has released official Qwen3.8 model weights, but they use the custom Qwen3.8-Max License rather than Apache 2.0. The precise description is “official open-weight release.” The hosted qwen3.8-max service remains a managed API product and is not itself a repository ID. Review the complete license instead of applying a family-wide “open source” label.
Methodology and update log
This update checked the QwenCloud release changelog, text-model selector, model-specific Marketplace, OpenAI-compatible Responses reference, Token Plan Individual and Team allowlists, Token Plan usage restrictions, billing FAQ, and Alibaba Cloud Model Studio model matrix on August 3, 2026. Prices and service limits were transcribed from the live model-specific QwenCloud page. No independent inference, benchmark, latency, tool, or multimodal result is claimed.
Update log: August 13, 2026 — added the official Qwen/Qwen3.8-2.4T-A95B and FP8 repositories, documented their text-only and thinking-only behavior, native and extended context limits, and custom license; separated the open checkpoint from the additional features of hosted qwen3.8-max; and updated qwen3.8-max-preview to retired with documented automatic routing to production.
Official sources
- QwenCloud model release changelog
- QwenCloud text-generation model selector
- QwenCloud Qwen3.8 Max Marketplace page
- QwenCloud OpenAI-compatible Responses reference
- Token Plan Individual overview
- Token Plan Team overview
- Token Plan usage FAQ
- QwenCloud billing FAQ
- Alibaba Cloud Model Studio text-model matrix
- Official Qwen3.8 model card
- Official Qwen3.8 FP8 repository
- Qwen3.8-Max License
- Qwen3.8 official announcement