I thought @ 0G_labs was just a plate, because I looked at the official website for a long time and didn't know what to say. I saw the progress of decentralized training in the following tweet today, which is the same as the technical directions of several decentralized training techniques I previously shared. It's just that their model should have a maximum scale of 100 billion, there are papers available https://arxiv.org/abs/2506.21263 No code. Deepmind's @ Ar-Douillard also reposted this paper. The performance described in the paper: -107B models can be trained on a 1Gbps slow network -Compared to vanilla AllReduce, the speed has increased by 357 times, -Superior to OpenDiLoCo (OOM>20 ° B) and CocktailSGD (over compression reduces convergence). -357 times faster than AllReduce on Qwen1.5-107B- -Can still maintain convergence at a compression rate of up to 1000 times The first major achievement in the field of decentralized training was a 2023 paper by the DeepMind team of @ Ar-Douillard https://arxiv.org/abs/2311.08105 ， https://arxiv.org/abs/2502.12996 They are mainly training models for heterogeneous low-speed Internet GPUs. At first, the models were relatively small. But it should have inspired many people. For example, the @ PrimeInterlect team and the @ gensynai team (initially deployed on the @ polkadot ecosystem, I don't know why they chose to deploy on ETH's L2 later). They mainly focus on the asynchronous nature of RL reinforcement learning for model post training. Among them, @ PrimeInterlect has open-source code implementation https://arxiv.org/abs/2407.07852 https://www.primeintellect.ai/blog/opendiloco https://((github.com))/PrimeIntellect-ai/OpenDiLoCo There is also @ tplr.ai (just discovered today) in this field, which was done on Bittensor's subnet (Bittensor used to feel like a plate, but now it seems that there is still an ecosystem doing things) https://www.tplr.ai/research https://www.tplr.ai/papers/templar_paper.pdf http://arxiv.org/abs/2505.23725 https://arxiv.org/abs/2505.23725 https://((github.com))/tplr-ai/CCLoco https://templarresearch. (substack.com)/p/ccloco-scaling-up-top-k-error-feedback There is also a team working in this direction @ NousResearch The team researching this direction can pay attention to it