Rednote Open-Sources “Relax,” Boosting Multimodal RL Training by 76%
Rednote has open-sourced a reinforcement-learning (RL) training engine, “Relax,” claiming up to a 76% end-to-end speedup for multimodal model training—an efficiency gain that could lower compute costs and shorten iteration cycles for Chinese AI teams racing to deploy agentic and multimodal applications in 2026.
The release, dated April 15, positions the lifestyle-content platform not only as an AI application owner but also as an infrastructure contributor. In early performance disclosures, Rednote said Relax achieved stable RL convergence across image, text, audio and video on Qwen3-Omni-30B, and outpaced ByteDance’s Volcano Engine veRL by 20% on a DAPO-Math task running on a two-machine, 16-GPU setup.
Market impact is immediate at the ecosystem level: by targeting RL’s bottlenecks—multimodal data throughput, fault tolerance in long multi-GPU runs, and scheduling inefficiency between rollout generation and model updates—Relax aims to make RL more production-like, where uptime and predictable throughput matter as much as peak FLOPS.
Targeting Token Bloat Reframes Multimodal RL Economics
Rednote tied its engineering rationale to “token explosion” and preprocessing overhead when images and audio-video enter the RL loop. Large raw files raise CPU preprocessing and transport costs, and multimodal encoders can clash with existing parallelization strategies—friction that tends to show up as unstable throughput and higher out-of-memory risk during long training jobs.
For investors and suppliers, the key implication is that multimodal RL is becoming a systems problem rather than a model-only problem. As more Chinese teams push agentic workflows (multi-turn reasoning, tool use, retrieval) into training, the winners will be those that can turn RL from an experimental pipeline into a repeatable manufacturing process—reducing wasted GPU time caused by synchronization stalls, retries and cluster-level restarts.
Separating Resources Shifts the Bottleneck From GPUs to Scheduling
Relax is built on Megatron-LM and SGLang, and uses a “co-design” approach to unify three constraints Rednote highlighted: full-modality data support, a service-oriented fault-tolerant architecture, and an asynchronous training pipeline.
The headline gain comes from moving beyond “Colocate” deployments—where training and inference share the same GPUs—toward a fully asynchronous design with finer-grained pipeline scheduling. Rednote reported that, versus a Colocate baseline, Relax’s fully async training delivered a 76% end-to-end acceleration. The company attributed the improvement to (1) streaming micro-batch scheduling that avoids long-tail blocking from global batch synchronization, and (2) resource separation that hides forward inference behind training time while eliminating sleep/wake overhead typical of shared-GPU setups.
That framing matters competitively: RL throughput is often limited not by raw GPU compute but by pipeline coordination across heterogeneous tasks. If Relax’s scheduling holds up in wider adoption, it could reduce the compute premium typically associated with RL fine-tuning—an advantage when cloud GPU availability and utilization efficiency are strategic constraints across China’s AI stack.
Hardening Fault Recovery Positions RL for Long-Run Production Use
Rednote packaged each RL role (Actor, Critic, Rollout and others) as an independent Ray Serve deployment with isolated fault domains, quotas and health checks. The design supports role-level failure isolation, differentiated recovery for stateless vs. stateful components, and independent scaling—so adding rollout inference replicas does not force coordinated resizing of the Critic cluster.
The company also described a distributed checkpoint service that can distribute updated weights to inference engines at low latency, aiming to restore service without reverting to disk checkpoints. In long-running jobs spanning large GPU fleets—where NCCL timeouts and hardware faults are routine—this architecture tries to reduce the “blast radius” of a single failure from a full restart to a localized recovery, protecting end-to-end training time.
Optimizing MoE Routing Narrows the Gap With Rival Frameworks
For mixture-of-experts (MoE) RL stability, Relax implements R3 (Rollout Routing Replay). On Qwen3-30B-A3B, Rednote said Relax with R3 reduced routing mismatch between training and inference by 38% while adding 1.9% overhead. By contrast, it said veRL’s R3 increased end-to-end time by 34%.
If reproducible, that delta speaks to a broader platform contest: RL engines are becoming differentiators for cloud AI offerings and large-scale in-house training. By open-sourcing Relax, Rednote is effectively pushing its systems choices into the developer toolchain—where adoption can translate into ecosystem influence, and where performance claims are validated in public by downstream users.
Related Coverage:
China's Instagram-like Xiaohongshu Expects Profit to Triple to $3 Billion in 2025
Xiaohongshu Challenges Big Tech AI With Open-Source, On-Premise Voice System