Fireworks' ARCv3 cuts RL weight-update payload by ~50% with lossless BF16 reconstruction
sophiamyang · x · 2026-09-24
Fireworks released ARCv3, a compressor for RL weight updates sent from trainers to rollout fleets. Across 1,000 production weight-update deltas it produced payloads 50% smaller than ARCv2 while losslessly reconstructing exact BF16 weights. Smaller transfers keep rollout fleets closer to the current policy, making cross-region training practical without one giant co-located cluster. Available via fireworks-delta-compression in the Fireworks Training API.
More from Infra
- GPU Demand Surges in Healthcare AI as Neoclouds Shift to Spot Pricing — Arindam_1729 · 2026-09-24
- DeskPilot: open source native Python desktop agent client for local LLMs with MCP and sandboxing — poofph · 2026-09-24
- Paper benchmarks skip gVisor; perf nerds eye memory overcommit and scheduler wins — sloppenheimer · 2026-09-24
- vLLM Tests Show TPUv7 Hits 700 tok/s/user, 56% Faster Than NVIDIA GB200 NVL72 — woosuk_k · 2026-09-24
- Theory: model 'nerfing' may come from mixed heterogeneous inference hardware, not intent — michellechen · 2026-09-24
- GKE 1.37 adds native scale-to-zero, ending reliance on KEDA setups — rseroter · 2026-09-24