🔥 Will Cocoon make a comeback?

⚫️ Back in August, we wrote that Cocoon’s public network had effectively stopped working and that its code hadn’t been updated in quite a while. Almost two months later, new commits have appeared in one of the project’s forks. Judging by the changes, the developer is trying to tackle one of the biggest challenges in running large AI models: limited GPU memory.

⚫️ For example, running GLM-5.3, one of the most powerful open-weight AI models, requires around 750 GB of VRAM. For comparison, an NVIDIA H100 — one of the most widely used data center GPUs for AI workloads — comes with just 80 GB.

⚫️ In other words, running a model like this normally requires several expensive GPUs at once. But what if independent GPU server operators could pool their compute and run the model together? That’s exactly the idea now being explored in Cocoon.

⚫️ Surprisingly, the work isn’t being done by Telegram, but by kdimentionaltree, one of the TON Core developers. So far, he has only managed to establish communication between virtual machines — distributed inference across multiple independent servers is not working yet.

⚫️ We don’t know whether this work will ever make it into a finished product, but it’s still pretty interesting to see a TON Core developer working on Cocoon. It could be another sign that TON Core and Telegram are starting to work more closely together.

@thedailyton