Open-Sourced Weight Teleportation Syncs RL Weights Across Data Centers in Under 0.23s

The gharik team has released and open-sourced a large-scale RL weight synchronization technique called "weight teleportation." It tackles the problem of pushing weight updates from trainer to sampler, which is typically huge in data volume, slow, and expensive. According to the announcement, even when the two sides sit in different data centers, sampler pause time can be kept under 0.23 seconds—internet-scale speed—breaking the traditional architectural assumption that trainer and sampler must live in the same data center. It's being hailed as a breakthrough on a key bottleneck for decentralized RL. The news spread the same day via reposts from Kyrannio, akiroai, and other accounts.

Confirmed

• The technique was proposed by the humansand team and released as open source by the gharik team, enabling high-speed weight sync between geographically separated trainers and samplers
• Key performance numbers: sampler pause time under 0.23 seconds, roughly 3… (the original post lists additional performance data)
• Addressing questions about cross-data-center latency, gharik relayed zianglih's explanation: the pause time only covers native decompression—decompression and patching done by the ASIC—and overlaps with local GPU work, so it is unaffected by distance or bandwidth between the two sides

Why it matters

• Traditional distributed RL architectures are locked into a single data center, and weight sync is the core weak point of decentralized RL; if the scheme works as described, it could substantially cut the cost and latency of RL training plus inference sampling deployed across data centers or even across regions

2026-11-03 ~ 2026-11-03 · 8 related posts

• [source] Weight teleportation open-sourced: cross-datacenter RL weight updates under 0.23s — gharik · 2026-11-03
• Team open-sources ultra-fast weight updates for RL, even across datacenters — akiroai · 2026-11-03
• Weight Teleportation makes cross-cluster RL weight updates faster than intra-cluster ones — Kyrannio · 2026-11-03
• [source] Sampler Pause Stays Under a Second Even Across Distant Datacenters — gharik · 2026-11-03
• RL weight-update pause time doesn't depend on datacenter distance, authors say — gharik · 2026-11-03
• Weight teleportation cuts RL sampler pauses below 0.23s at internet speeds — gharik · 2026-11-03
• humans& details GPU-accelerated Weight Teleportation: lossless compression for instant RL weight sync — niloofar_mire · 2026-11-03

1 near-duplicate retellings: gharik