Actual networks

In general, for game analysis and long calculation time per move, the largest network compatible with your hardware is recommended. In use cases with very low calculation time per move or slow hardware, a smaller network might be a better choice.

T1 networks above are contributed by masterkni6. The larger 768x15 network is comparable in architecture with networks in current training run1.

If you’re getting out of memory errors when using large networks on GPU, pick the next best network in the list or try adding --backend-opts=max_batch=256 to LC0 command (or UCI option: BackendOptions: max_batch=256), default: 1024. This will reduce GPU memory usage without any negative impact on playing strength. With the cuDNN backend you can also try --backend-opts=custom_winograd=false or as a UCI option: BackendOptions: custom_winograd=false.

Note for DirectX12 and OpenCL backend users: The format of the networks in the list above is not supported. However, you can download and use the LC0 ONNX-DML version instead, see the included README file for instructions on how to get the directml.dll that can’t be included in the package for licensing reasons. Alternatively you can use older networks such as the last T78 512x40 network 782344 or the last T60 384x30 network 611246.

Network Lists

Listed for completeness, includes networks from older training runs. Some download links might be outdated.

In each section, the nets are listed roughly in descending order of strength. Some may be too close to tell apart.

30 blocks x 384 filters:

24 blocks x 320 filters:

20 blocks x 256 filters:

15/16 blocks x 192 filters:

10 blocks x 128 filters:

Assorted sizes:

If you still have questions, check the Discord channels. Be sure to specify your hardware and use case so the helpful regulars know what to recommend.