Fireworks releases ARCv3 for smaller lossless RL weight updates
Fireworks says ARCv3 reduces BF16 weight-update delta payloads to an average 0.19% of original weight size, compared with 0.36% for ARCv2, while reconstructing the trainer’s exact weights. It is the default for Fireworks Trainer SDK rollouts; teams bringing their own trainer can integrate it through the fireworks-delta-compression package.
SessionWatcher editorial · Published · Updated · Source announcement: 2026-09-23
What ARCv3 changes
Fireworks says ARCv3 compresses BF16 weight updates sent from a trainer to machines generating reinforcement-learning rollouts. In its benchmark of 1,000 production RL weight-update deltas, the average compressed payload was 0.19% of original BF16 weight size, compared with 0.36% for ARCv2. Fireworks describes reconstruction as lossless: the rollout side recovers weights bit-for-bit identical to the trainer’s.
Sources: Every byte counts: ARCv3 and the case for cross-region RL
Who can use it and how
Teams using the Fireworks Trainer SDK for rollouts do not need to change code; ARCv3 is already the default. Teams using their own trainer with Fireworks rollouts can integrate the fireworks-delta-compression package. The post gives the installation command as pip install fireworks-delta-compression and describes selecting the arc_v3 compression format in code.
Sources: Every byte counts: ARCv3 and the case for cross-region RL
What the result means
The compression is for weight updates, not a general model-file or inference-cost reduction. Fireworks says smaller transfers can help distributed rollout fleets stay closer to the current policy and use compute across regions. Its benchmark compares payload sizes on tested BF16 tensors; it does not establish a general performance result for other workloads.
Sources: Every byte counts: ARCv3 and the case for cross-region RL
AI assisted reporting, checked against the linked official sources. Source pages checked 2026-09-30. Editorial process and corrections.