Fireworks says its new ARCv3 compressor produced an average payload of 0.19% of original BF16 weight size across 1,000 production RL weight-update deltas. The same company benchmark lists 0.36% for ARCv2, 0.35% for PULSESync, and 0.67% for a draft Hugging Face TRL implementation. It also says roughly 2% of BF16 weights changed between consecutive checkpoints in the workloads examined, and that ARCv3 reconstructs the trainer’s exact BF16 weights losslessly.
The mechanism is to omit unchanged values, encode the common mantissa-only changes with only the relevant bits, and separately pack rarer exponent or sign changes. ARCv3 additionally tries several general-purpose codecs per tensor and retains the smallest output. This matters now because frontier-scale RL repeatedly distributes policy updates to rollout capacity, while access to a single huge co-located cluster remains constrained and expensive.
The implication is that communication engineering can expand the usable geography of RL compute rather than merely reduce storage or egress costs. If update propagation falls behind, rollout workers act on an older policy and distributed training loses its advantage; smaller exact deltas directly attack that synchronization gap. The evidence boundary is that the reported results are Fireworks’ own benchmark on its production deltas, not an independent evaluation across models, networks, or training regimes. Watch whether independent workloads reproduce the advantage, whether decompression and codec-selection CPU remain negligible at larger scales, and whether lower transfer volume yields measurable end-to-end RL quality or throughput gains.