Owned by AlibabaTongyi QianwenThe team released on August 12 Qwen3.8 2.4T A95B, which is the open source weighted version of Qwen3.8 Max, has2.4 trillion total parameters, 95 billion active parameters, using Sparse Mixed Experts (SMoE) architecture, is currently one of the open source weight models with the largest parameter scale. The API is priced at $2 (input) / $6 (output) per million Tokens, and supports a context window of 1 million Tokens.
core competencies
The release of Qwen3.8 2.4T A95B marks another major breakthrough in China’s AI open source ecosystem:
- 2.4T parameter scale: The total parameters are 2.4 trillion and the active parameters are 95 billion. It is one of the largest parameter scales among the current open source models.
- Sparse MoE architecture: Only 95 billion parameters are activated for each inference, controlling inference costs while maintaining top performance
- Millions of contexts: Extra-long context window of 1,048,576 Tokens to handle complete code bases and large documents
- Open source weight: As an open source version of Qwen3.8 Max, developers can freely download, fine-tune and deploy
- Multi-provider support: There are already 7 providers on OpenRouter, including Together AI, Fireworks, etc.
User experience and limitations
The Qwen3.8 2.4T A95B performs well on programming, mathematical reasoning, and multi-language tasks. As an open source weight model, it provides enterprises and developers with completely independent and controllable deployment options - no need to rely on third-party APIs and can be deployed privately. The parameter scale of 2.4T means that it is capable of handling extremely complex inference tasks, and the SMoE architecture makes the inference cost controllable.
But be warned: API pricing of $2/$6 is on the high side in an open source model - by comparison DeepSeek V4 Pro is only $0.435/$0.87. In addition, downloading the complete 2.4T model weight requires extremely high hardware configuration (at least multiple A100/H100), and ordinary developers may only be able to call it through API calls. The maximum output is only 52,429 Token, compared to DeepSeek There is a clear gap between 384K.
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 9.3 / 10 | 2.4T parameters + 1M context + SMoE architecture, the top configuration in the open source model |
| 易用性 | 7.5 / 10 | The threshold for downloading open source weights is high (requires multi-card cluster), API calls are more practical but the price is high |
| Cost-effectiveness | 7.0 / 10 | $2/$6 is relatively high among open source flagships, but as an open source weight model, it can be deployed privately and the long-term cost is controllable. |
| 中文支持 | 9.5 / 10 | AliTongyi QianwenProduced by a team with top-notch Chinese understanding and generation capabilities in the industry |
| 输出质量 | 9.0 / 10 | Inheriting the reasoning ability of Qwen3.8 Max, it performs well in programming and mathematics tasks |
Overall rating: 8.5/10
Qwen3.8 2.4T A95B is a milestone in the open source AI community - the open source weight model of 2.4T parameters allows enterprises and research institutions to have autonomy in top AI capabilities. Although API pricing is on the high side, for scenarios that require privatized deployment and high data security requirements, this is currently one of the best options.
To learn more about AI model reviews, please visit AI Dash — Discover the best AI tools.