humans&, the AI lab, says it has cut the freeze that interrupts a model while it learns from its own answers to less than 0.23 seconds. Engineer Ziang Li describes the method, which the lab calls weight teleportation, on the humans& blog, and the code is open source inside the Miles and SGLang projects.
The pause matters because of how this kind of training works. In reinforcement learning, one group of GPUs (the trainer) improves the model, while another group (the sampler) uses the latest version to produce answers that get scored. After each training step the sampler needs the new version, and while it swaps, it generates nothing. Idle accelerators cost the same per hour as busy ones, so a long freeze on every step is money burned. The humans& post also argues the two sides should hold matching weights as closely as possible, which means the swap cannot be skipped.
The obvious way to send an update is to ship the whole model. For the lab’s test model, an NVFP4 build of GLM-5.3, that is 445 GB. Li’s team instead sends only what changed. A single training step touches a sliver of the weights (the post’s test changes 0.085% of the checkpoint), so the trainer compares old and new versions bit by bit and produces a map that is almost entirely zeros. Zeros compress extremely well.
Sending only changes is an existing idea, and the post credits the Miles project’s earlier disk-based version. In that version the compressed payload was only 0.876 GB, but the gain disappeared into compression time, writing a checkpoint to disk, and reloading the entire model afterward.
The new design moves all of that onto the GPUs. Nvidia’s Blackwell chips carry dedicated decompression hardware that handles formats such as LZ4 and Snappy. LZ4 alone produced a 5.36 GB payload on the test update, about six times the CPU-based Zstd result. So the lab stacked two layers: Snappy first, then Zstd on top, which brought the transfer down to 2.65 GB, about 0.6% of the model. The outer layer is unpacked by the CPU while the sampler keeps working. Only the last step, where the decompression hardware streams the changes into the live weights one layer at a time, stops generation, and the post puts that at 219.6 milliseconds.
The wider timings are modest too. Preparing and compressing on the trainer takes 0.789 seconds, the sampler side needs 2.355 seconds to fetch and stage the payload without stopping, and the lab reports under four seconds from training step to updated sampler. The trainer in the test is 64 B300 GPUs across eight nodes.
Every number here comes from humans& testing its own workload on its own hardware, and the post offers no outside replication. Its claim that the method beats even the fastest direct GPU-to-GPU links inside a cluster, and its statement that it foresees no further work on weight sync, are the lab’s own assertions. The approach also leans on Blackwell’s decompression hardware, so teams on older chips would not see the same result.
Li adds a second use. Because a delta is tiny and applies with one API call and the same sub-0.23-second pause, a serving system could swap fine-tuned variants of a model in and out, which the post pitches as a rival to LoRA adapters that does not limit how much the update can change. For anyone running reinforcement learning on Blackwell with SGLang, the first thing to measure is how much of their current step time is spent waiting on exactly this swap.
Reported by the humans& blog (humansand.ai), in a post by Ziang Li; the file carries no publication date.
