Qwen 3.8 Flash Next 125B Model can now Work at 100T/s Speed on Consumer GPU

Artificial intelligence development community has taken great language models more important step in consumer hardware. A new project called Strata is able to run 125 billion parameters of Alibaba’s powerful Qwen 3.8 Flash Next model with impressive performance on consumer GPUs such as RTX 4090.

The success achieved under the project is as follows:

-125 billion parameter model It can work on a single consumer GPU like RTX 4090

• 100 tokens/second speed output production is provided, which is enough speed for real-time applications
• Strata’s optimization techniques combine model compression and efficient memory management
• Thanks to open source solution, researchers and developer ring are able to experience with these models easily
• The project on GitHub invites the participation and development of a wider community

This development creates a great opportunity for many researchers and developers that can not access professional data center infrastructure. The model run operations that require millions of dollars are now made with a price tag under the handlib.

The success of Strata shows how effective AI community’s democracy efforts and optimization techniques are. Many larger models in the future can be expected to operate in consumer hardware.