Fireworks cuts RL delta size in half, makes cross-region training practical

The bottleneck in frontier model training isn't the GPU supply — it's the network. Fireworks is launching ARCv3, a new compressor for reinforcement-learning weight updates that cuts the payload nearly 50% compared with its predecessor. The release targets the least-discussed constraint in RL at scale: not the shortage of H100, but the pipes that move the updated model from the trainer to the fleet of machines generating rollouts. When those pipes can't keep up, the only option is a massive, co-located cluster — expensive, scarce, and a barrier to smaller players.

Fireworks' earlier post, "Frontier RL is cheaper than you think," showed that in the RL runs it examined, only about 2% of BF16 weights change between successive checkpoints. Instead of pushing 1 terabyte to every rollout machine at each step, Fireworks sends a compressed delta — a compact description of what changed — and reconstructs the model on the receiving end. That approach is what makes synchronized training across three or four regions feasible without dedicated wide-area links. Delta size determines whether the fleet stays close to the current policy or drifts stale while the trainer pulls ahead.

ARCv3 brings the average payload down to 0.19% of the original weight size, from 0.36% in the previous version — lower is better. The compression is lossless: the model reconstructed on the rollout side is bit-for-bit identical to what the trainer produced. Each trainer shard uploads its portion of the compressed delta to object storage (S3) in parallel; the Fireworks API signals every region that an update is available; each region pulls the relevant shards and reconstructs locally. The trainer never speaks directly to the rollout fleet, which is what lets a distributed inference cluster stay in sync over ordinary network links.

The improvement comes from a statistical asymmetry Fireworks observed in production. Of the roughly 2% of weights that change in a given step, the vast majority undergo only a mantissa shift, while the exponent and sign remain fixed. Updates touching the exponent are rare; sign flips are rarer still. In other words, most updates are small nudges, not jumps. ARCv3 encodes that asymmetry directly: unchanged values are omitted from the payload, and mantissa-only changes contribute only their mantissa bits.

The compressor is available now through the fireworks-delta-compression package, for teams using their own trainer with Fireworks rollouts. For Israeli groups pushing frontier-scale RL — whether in startups, research labs, or large organizations — the practical upshot is clear: the inference fleet can be deployed in whatever regions are available and cost-effective, without locking into a single expensive cluster, while still keeping rollouts updated at the trainer's pace.