Qwen 3.5 122B variant runs at 28.45 tokens per second
A compact version of Qwen 3.5 122B was released for local use on an system. The model has 122 billion parameters in total, but only 10 billion are active during each step.
It takes 60.70 GiB of memory and generates 28.45 . It processes prompts at 353.3 .
Its measured KLD against the original model is 0.100716. This was the creator's first model , and a version without the Heretic changes may follow when enough is available.
Key points
- The model has 122 billion total parameters and activates 10 billion per processing step.
- It uses 60.70 GiB of memory on hardware.
- It generates 28.45 .
- reaches 353.3 .
- A version without the Heretic changes may be released if allows.