Instructions to use EschaLabs/escha-runtime-qwen3moe with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use EschaLabs/escha-runtime-qwen3moe with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir escha-runtime-qwen3moe EschaLabs/escha-runtime-qwen3moe
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Windows support?
I have downloaded everything and was installing dependencies and saw pip install ./sglang/escha-*.whl
the whl file is escha-1.0.2+qwen3moe-cp312-cp312-manylinux_2_28_x86_64.whl.
All that work was for nothing? Is there a way to run on Windows os?
First of all, thanks for trying out the model!
To answer your Windows OS support question:
The weights are fine. That's the 12GB, and it's the same files for every engine we ship. Only the wheel is Linux-only (it's a compiled CUDA extension plus our sglang fork).
For Windows the answer is WSL2. wsl --install -d Ubuntu-24.04 (you want 24.04, it has python 3.12), then follow sglang/INSTALL.md as written. CUDA in WSL2 goes through your existing Windows driver, so there's nothing extra to install.
Two things that'll bite you: WSL only takes ~half your RAM by default and load peaks higher than you'd expect, so put memory=24GB under [wsl2] in .wslconfig. And copy the weights over to the linux side rather than loading from /mnt/c, that mount is slow.
I haven't run it on WSL2 myself so I can't promise it's clean, but nothing in the stack should care. Tell me how it goes.
Native Windows isn't happening for now, upstream sglang doesn't build there.