Qwen3.8-2.4T-A95B is the open-source release of Qwen's latest flagship, launched August 2026. Its sparse MoE architecture holds 2.4T total parameters with ~95B activated per step, paired with hybrid attention and a 1M token context window. Key benchmarks: GPQA Diamond 92.6, PaperBench 93.0, OSWorld 86.1, BabyVision 82.0. Ranked 4th on CodeArena.
Inference Service Provider
The inference service provider for qwen3.8-2.4t-a95b is Alibaba Cloud Model Studio.
Model Capabilities
| Capability | Support | Capability | Support |
|---|---|---|---|
Input Modality | Text | Output Modality | Text |
Model Experience | Supported | Function Calling | Supported |
Structured Outputs | Supported | Web Search | Supported |
Prefix Completion | Supported | Context Caching | Supported |
Batch Inference | Unsupported | Fine-tuning | Unsupported |
Context Limits
Parameter | Value | Parameter | Value |
|---|---|---|---|
Max Input Length | 991808 | Max Output Length | 131072 |
Context Length | 1000000 | Max Input Length (Thinking Mode) | 983616 |
Max Output Length (Thinking Mode) | 131072 | Max Chain-of-Thought Length | 131072 |
Pricing
This page only shows the original pricing for model API calls, excluding any limited-time promotions. Visit Model Studio Console for promotional offers.
China (Beijing)
| Billing Item | Price (USD) | Unit |
|---|---|---|
Input | 1.65 | Per 1M tokens |
Output | 4.951 | Per 1M tokens |
Input (Cache Hit) | 0.206 | Per 1M tokens |
Explicit Cache Creation | 2.063 | Per 1M tokens |
Explicit Cache Hit | 0.137 | Per 1M tokens |
Singapore
Scope: International
| Billing Item | Price (USD) | Unit |
|---|---|---|
Input | 2 | Per 1M tokens |
Output | 6 | Per 1M tokens |
Input (Cache Hit) | 0.25 | Per 1M tokens |
Explicit Cache Creation | 2.5 | Per 1M tokens |
Explicit Cache Hit | 0.17 | Per 1M tokens |
Rate Limits
China (Beijing)
This region uses dynamic rate limiting. The TPM limit is tiered by your monthly Model Studio spend, and the RPM limit is high enough that normal use does not trigger throttling. For the tier values, see Dynamic rate limiting.
Singapore
Scope: International
This region uses dynamic rate limiting. The TPM limit is tiered by your monthly Model Studio spend, and the RPM limit is high enough that normal use does not trigger throttling. For the tier values, see Dynamic rate limiting.