Skip to main content
Language modelActive

Qwen 3.8 2.4T A95B

Alibaba

Released
August 12, 2026
Data date
September 8, 2026
Speed
39.8 tokens/s
Output throughput
Price
$2 · $6
Input · Output / 1M tokens
Input
Text
Output
Text

Qwen 3.8 2.4T A95B is the open-weight checkpoint behind the Qwen3.8-Max serving release. Its repository contains FP8-quantized weights for a hybrid Mixture-of-Experts model with 2.4 trillion total parameters and 95 billion active parameters. The architecture combines Gated DeltaNet and Gated Attention.

The weights are text-only and require Thinking by default. Native context reaches 262,144 tokens and can be extended to roughly 1,010,000 tokens. The hosted Qwen 3.8 Max is a different configuration. Its model card says it adds vision input, switchable Thinking, a default 1M context, and built-in tools.

Specifications and access

SpecificationValue and source
Weights
Open weightsSource
Access
API, LocalSource
Parameters
2.4TSource
Active parameters
95BSource
Context window
262,144 tokensSource
Weights repository
Qwen/Qwen3.8-2.4T-A95B-FP8Source
License
qwen3.8-max (The model card declares a non-SPDX custom/other license; inspect the linked LICENSE before publishing a generic license badge.)Source
Architecture
hybrid Gated DeltaNet + Gated Attention sparse MoESource
Input
TextSource
Output
TextSource
Reasoning mode
thinking by default; switchable non-thinking/instruct modeSource
Tool calling
YesSource
Structured output
YesSource
Native context window
262,144 tokensSource
Maximum context window
1,010,000 tokensSource
Parameters
2,400 BSource
Active parameters
95 BSource
Repository created
2026-08-08 (Repository creation date is a precise publication proxy, but may differ from the public announcement date or an individual checkpoint's first availability.)Source