Is moonshot AI team going to open source the full fp16 precision model?
It doesnt make sense to open source only the 4 bit quantized version of the model and keep the actual fp 16 locked away.
Unless this model does not have a 8 bit or 16 bit version at all.
Was it trained natively in 4 bit only?
Not necessarily. It could be native 4-bit training rather than post-training quantization. If you look at models like DeepSeek-v4-pro、gpt-oss-20b/120b (which use MXFP4), the 4-bit precision is part of the training itself, meaning the 4-bit version is the model.
read the technical report bro, it is native mxfp4.
If memory serves me correct the four-bit quantization only applies to the moe expert layers, not the entire model's layers. The GPT-OSS models are also MXFP4 and this seems to be primarily done to allow efficient serving of the model (maintaining fast t/s and near indistinguishable performance compared to 8bit or 16bit versions while consuming less resources. This matters a lot whether you're running the model on your own hardware or if you're an inference provider or even a regular user calling the API to use the model because this means lower costs and faster generation speeds for everyone.