Fireworks ARCv3 halves RL weight update payloads for cross-region training

Fireworks AI released ARCv3, a new compressor for reinforcement learning weight updates, on September 23, 2026. The compressor reduces the size of BF16 weight deltas sent from the trainer to rollout machines, with an average payload of 0.19% of the original BF16 weight size, down from 0.36% with ARCv2. The compression is lossless, meaning the reconstructed model on the rollout side is bit-for-bit identical to the trainer's output.

How ARCv3 works

ARCv3 exploits an asymmetry observed in production RL training sessions. Of the roughly 2% of BF16 weights that change between consecutive checkpoints, most see only a mantissa shift, while the exponent and sign remain unchanged. Updates that touch the exponent are rare, and sign flips are rarer still. ARCv3 encodes this by omitting unchanged weight values, contributing only mantissa bits for common changes, and packing rare exponent or sign updates as separate blocks. The rollout side applies both streams to the previous checkpoint and checksums the result against the trainer's original.

For each tensor, ARCv3 runs several general-purpose compression algorithms in parallel and keeps the smallest output. This costs extra CPU, but trainer machines typically have spare CPU cycles while GPUs handle training.

Benchmark results

Fireworks benchmarked ARCv3 against three alternatives on 1,000 real weight-update deltas from production RL workloads, all in BF16 and lossless. The compressed-to-original ratio was 0.19% for ARCv3, 0.35% for PULSESync, 0.36% for ARCv2, and 0.67% for Hugging Face TRL's draft implementation. ARCv3 produced payloads roughly half the size of the next-best approach.

Using ARCv3

For teams using the Fireworks Trainer SDK for rollouts, ARCv3 is already the default, requiring no code changes. Teams bringing their own trainer can integrate ARCv3 through the fireworks-delta-compression package, available in the Fireworks Training API. The package provides delta_compress and delta_decompress functions that accept BF16 tensors and return compressed diffs with checksums.

Why it matters

Smaller deltas help rollout fleets update faster and stay closer to the current policy, making it more practical to use compute across regions without a single co-located cluster. Fireworks' approach uses shared object storage: each trainer shard uploads its slice of the compressed delta to S3 in parallel, then the API signals all regions to pull and reconstruct the model locally. The trainer never talks directly to the rollout fleet, allowing globally distributed inference clusters to stay in sync over ordinary network links.