# Baseten — Optimizing delta weight syncs for managed RL rollouts

- Company: Baseten (baseten.co)
- Announced: 2026-09-08
- Category: infrastructure-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.baseten.co/blog/optimizing-delta-weight-syncs-for-managed-rollouts
- Record: https://forck.live/items/32773-optimizing-delta-weight-syncs-for-managed-rl-rollouts
- Subject: Baseten
- Models affected: GLM-5.3

Baseten says it can sync delta weights to frontier models such as GLM-5.3 across global, independent clusters in under 40 seconds, with only 6 seconds of request pause. The company's managed rollouts let AI labs keep their trainer in their own environment while Baseten runs the rollouts on the same infrastructure that serves inference for Cursor, Hubspot, Lovable and Notion. The trainer publishes full checkpoints and small deltas to object storage, and each rollout cluster independently discovers, downloads and applies the new version without blocking the trainer or other clusters. Deltas are byte-level XORs of serialized weights from consecutive policies; on Baseten's GLM-5.3 fixture, 716.56 GiB of logical XOR data compressed to 1.55 GiB with zstd. Rather than decompress and scan that full output, Baseten's custom decoder reads zstd's literals and sequence commands directly, recording only non-zero literals and skipping match ranges that contain only zeroes. The decoder then groups changed bytes into tensor elements, fetches them from CPU memory and applies XOR to produce absolute new values for the GPUs.

## Evidence

Verbatim from https://www.baseten.co/blog/optimizing-delta-weight-syncs-for-managed-rollouts:

> Baseten supports delta weight syncs to leading frontier models like GLM-5.3 across global, independent clusters in under 40 seconds, with only 6 seconds of request pause. AI labs and frontier research teams use our managed rollouts offering to run their frontier-scale RL. They control the trainer and we run the rollouts on the same infrastructure that powers inference for Cursor, Hubspot, Lovable, and Notion.
> 
> A trainer runs in the customer's environment. It publishes full checkpoints and small deltas to object storage. A Baseten deployment leveraging a global pool of capacity watches the same stream, reconstructs the requested policy, updates GPUs across our fleet, and keeps serving. We use the Baseten Delivery Network for the large object transfers so each cluster only downloads model weights once which minimizes repeated public-cloud egress costs. See Stephen Day’s explanation of the Baseten Delivery Network here .
> 
> Utilizing a global fleet of clusters
> 
> Many very fast weight-update systems assume the trainer and rollout GPUs are connected by the same high-bandwidth network. We do not.

---

Record: https://forck.live/items/32773-optimizing-delta-weight-syncs-for-managed-rl-rollouts
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md
