How much does the Qwen 3.8 Max API cost?

2 dollars per million input tokens and 6 dollars per million output tokens, with cache reads from 0.17 dollars per million. Rate limits at launch are 2 million tokens per minute and 15,000 requests per minute.

Can I use Qwen 3.8 with tools and structured outputs?

Yes. Function calling, structured outputs, batch jobs, prefix completion, and fine tuning are supported, plus built in tools: code interpreter, web search, web extractor, and image search.

Can I self host Qwen 3.8?

The Max weights are public since August 12, 2026, but at roughly 1.2 terabytes of memory at 4 bit it is still datacenter hardware. The Qwen 3.8 27B from the same wave is the realistic self host target: Q4_K_M measures 17.1 GB, and IQ4_XS at 15.7 GB is the largest quant that fits a 16 GB card whole.

How to Run Qwen 3.8 Max: Every Option That Actually Works

Q: How do I access Qwen 3.8 Max?

Through Alibaba's DashScope platform. Create an API key in the DashScope console, then call the OpenAI compatible endpoint with any OpenAI SDK. The API also accepts Anthropic format requests, so Claude style clients work by swapping the base URL and key.

Q: Is Qwen 3.8 uncensored?

No, the hosted model carries the usual alignment and refusals. Uncensored community builds only become possible once the weights are public. Today the strongest uncensored Qwen options are abliterated builds of Qwen 3.6, which run locally on consumer GPUs.

Update, August 14, 2026: the wait is over. The 2.4T Max checkpoint went public on Hugging Face on August 12 as Qwen/Qwen3.8-2.4T-A95B under Alibaba's own qwen3.8-max terms, which is not Apache 2.0, and the consumer sized Qwen 3.8 27B followed on August 13 under Apache 2.0. Real GGUF sizes, the VRAM math and the chat template trap are in How to Run Qwen 3.8 27B Locally. Everything below is the pre drop picture, kept for context.

Qwen 3.8 Max went generally available on August 3, 2026. This is the practical guide: every way to actually use it today, with working code, real prices, and no pretending. If you want the specs and benchmark story instead, start with the explainer.

Option 1: The DashScope API (OpenAI Compatible)

The primary route is Alibaba's DashScope platform. Create an API key in the DashScope console, then point any OpenAI SDK at the compatible endpoint:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_DASHSCOPE_API_KEY",
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

resp = client.chat.completions.create(
    model="qwen3.8-max",  # check the console model list for the exact id
    messages=[{"role": "user", "content": "Summarize this repo's architecture."}],
)
print(resp.choices[0].message.content)

Streaming, function calling, structured outputs, and vision input all work through this endpoint. The built in tools (code interpreter, web search, web extractor, image search) are enabled per request via the tools parameter.

Pricing: $2 per million input tokens, $6 per million output tokens, cache reads from $0.17. Rate limits: 2M tokens per minute, 15K requests per minute at launch. Batch jobs and fine tuning are available from day one.

Option 2: The Anthropic Compatible Endpoint

New for this generation: the API also speaks the Anthropic request format. Tools built for Claude (including many coding agents) can point at Qwen 3.8 Max by swapping the base URL and API key, no request rewriting needed. If your agent stack is Claude shaped, this is the lowest friction way to try it.

Option 3: Routers and Aggregators

OpenRouter and the other aggregators typically list Qwen flagships within days of general availability. Until the weights are public there is only one upstream (Alibaba), so pricing tracks the official API and every request routes to the same place. The value is the one line model switch, not a better deal.

Option 4: Locally (Read This Before You Try)

The Max does not run on your hardware. Roughly 2.4 trillion parameters means ~1.2 TB of weights at 4 bit, before any KV cache for the 1M context. The full math lives in Can You Run Qwen 3.8 Locally?

The local path that IS real, and it is real now: Qwen 3.8 27B came out on August 13, 2026 under Apache 2.0. Q4_K_M measures 17.1 GB, so a 24 GB card holds it comfortably and a 16 GB card wants IQ4_XS at 15.7 GB instead. It is one click in Locally Uncensored's Model Manager, and the full setup, including the chat template trap that makes it look broken, is in How to Run Qwen 3.8 27B Locally. Qwen 3.6 is still the better answer below 12 GB.

Which Option for Which Job

You want	Use	Why
Max quality, production API	DashScope, OpenAI compatible	First party, full tool support, cheapest cache pricing
Claude shaped agent stack	Anthropic compatible endpoint	No request rewriting
Quick comparison against other models	OpenRouter	One line model switch
Privacy, zero refusals, free	Local models via LU	Qwen 3.6 today, Qwen 3.8 27B after the drop

What About Uncensored Use?

The hosted Qwen 3.8 Max carries the usual alignment and refusals, and API providers log by policy. If uncensored and private is the point, that is what local models are for: the abliterated builds of Qwen 3.6 are among the strongest uncensored models of 2026, and the moment Qwen 3.8 27B weights are public, the abliteration pipeline gets to work on a new generation. We will cover those builds when they exist.

FAQ

How do I access Qwen 3.8 Max?

DashScope API key, then the OpenAI compatible endpoint with any OpenAI SDK. An Anthropic compatible endpoint exists for Claude style clients.

How much does it cost?

$2 / M input, $6 / M output, cache reads from $0.17. Rate limits 2M tokens and 15K requests per minute.

Can I use tools and structured outputs?

Yes: function calling, structured outputs, batches, prefix completion, fine tuning, plus built in code interpreter, web search, web extractor, and image search.

Can I self host it?

The Max: no, ~1.2 TB at 4 bit. The 27B from the same drop: yes, ~16 GB at 4 bit. Full breakdown here.

Is Qwen 3.8 uncensored?

Hosted: no. Local uncensored builds become possible once weights are public; today's answer is abliterated Qwen 3.6.

Getting Started

While the API answers your hardest questions, a private, uncensored model can be running on your own GPU in the next five minutes:

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
# Windows: setup.bat | Linux: ./setup.sh

Or download the installer from the releases page, open the Model Manager, and pick a Qwen that fits your VRAM.

Locally Uncensored is AGPL-3.0 licensed and free to use. Built by PurpleDoubleD.