Qwen3.8 Max preview early access: 2.4T params and always-on thinking

Qwen ambassadors got early access to Qwen3.8 Max preview, and mine arrived. Qwen3.8 Max is a multimodal model announced at WAIC in Shanghai on July 19, 2026, handling images, video, and documents. The official X account put the total parameter count at 2.4 trillion (2.4T), the first Qwen model past 1T. No benchmarks have been published yet.

Japanese proofreading and agent-style testing will come in a separate post, following the same flow as my function-calling agent test with Qwen3.7 Plus via ModelScope.

First calls

I pointed an OpenAI-compatible SDK at the model. Connection details live in environment variables.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["QWEN38_API_KEY"],
    base_url=os.environ["QWEN38_BASE_URL"],
)

resp = client.chat.completions.create(
    model=os.environ["QWEN38_MODEL"],
    messages=[{"role": "user", "content": "こんにちは"}],
)
print(resp.choices[0].message.content)

Item	Detail
Connection	OpenAI-compatible API (Alibaba Cloud Model Studio)
Model	qwen3.8-max-preview
Client	Local Python (openai package)
Parameters	Defaults (temperature not set)

model field: qwen3.8-max-preview
reasoning_tokens: 124

The model field in the response comes back as qwen3.8-max-preview, so generation is really happening. usage includes reasoning_tokens, so this is a model that thinks internally before answering. models.list() doesn’t show any 3.8 models, though. The model isn’t in the listing, yet chat.completions went through.

I threw just two quick prompts at it. A simple multiplication (12345 × 6789) came back correct in about 6 seconds with 242 reasoning_tokens. A self-referential question — what changed between 3.7 and 3.8 in you? — got only “unknown” as the answer, with 980 reasoning_tokens and about 31 seconds.

Thinking is always on and reasoning_effort is fixed at xhigh

Under the early-access spec, enable_thinking=False does not disable thinking; it stays on. That’s why even the simple multiplication carried reasoning_tokens. The default reasoning_effort is xhigh, and the official benchmarks are measured at that setting. Change it and your numbers are no longer comparable to the official results.

preserve_thinking is also enabled by default; when enabled, the reasoning_content in the conversation history must be sent back intact with the next request. Concatenating reasoning_content into content is not supported. Coding and multimodal use are recommended to run in separate conversations rather than mixed in one.

These are preview-stage rules and may change at official release.

Active parameter count and MoE layout are not public

The 2.4T figure is total parameters; only a fraction of that should be used per token at inference. Qwen’s Max series has been MoE (Mixture of Experts) so far, with total capacity and active parameters treated as separate numbers.

Qwen3-Next’s gated attention rebuilt the attention side into a hybrid of Gated DeltaNet and Gated Attention on top of the MoE FFN. But Qwen3-Next and 3.8 Max preview are different models, and there is no public information on whether 3.8 inherits that design. Active parameter count and MoE layout are unknown for now.