Deel Data Science Team Runs RL Experiments in Hours, Boosts GPU Efficiency 5x

The iteration speed of reinforcement learning (RL) experiments has long been a bottleneck limiting the efficiency of AI teams. Recently, the data science team at fintech company Deel used Together's post-training platform to go from environment setup to running real RL experiments in just a few hours, and ran up to five experiments simultaneously on the same GPU capacity. This means effective GPU utilization increased several-fold, and the iteration loop accelerated significantly.

From Setup to Experiments: A Leap Within Hours

According to information shared by the Deel team on social media, they used Together's post-training service to complete the entire process from initial setup to actual RL experiments in a "couple of hours." For anyone familiar with reinforcement learning engineering, this speed is impressive—traditionally, setting up an RL training environment, debugging distributed configurations, and ensuring the reward function correctly interfaces with the training loop often takes days or even weeks.

Together's post-training platform has clearly done substantial work at the abstraction layer, packaging environment configuration, compute scheduling, and experiment management into more usable interfaces. The Deel team did not disclose specific technical details, but judging from the results, this "out-of-the-box" experience lowers the barrier to starting RL experiments.

Post-training in the LLM context refers to the training phase after a base model's pre-training is complete, typically including supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and the recently popular reinforcement learning with verifiable rewards (RLVR). Unlike pre-training, which requires massive unlabeled corpora, post-training relies on high-quality preference data or task environments that can be automatically scored, placing higher engineering demands on environment configuration, reward function implementation, and the training loop. A so-called "post-training platform" encapsulates the distributed training frameworks, environment management, experiment tracking, and compute scheduling in the above workflow as a managed service, so users only need to provide data and reward logic to launch experiments. This is precisely the precondition that enabled the Deel team to run RL experiments within hours—the platform absorbs the engineering complexity that would otherwise require weeks of work by an ML infra team.

Five Experiments on the Same GPU Capacity: The Multiplicative Effect of Efficiency

More noteworthy is the second data point: the Deel team subsequently ran up to five experiments simultaneously on the same GPU capacity. In the traditional model, an RL experiment often monopolizes a set of GPUs, and training jobs queuing for resources is the norm. Through more efficient scheduling and resource reuse, however, a single GPU pool can support multiple experiments running in parallel.

The direct benefit is an increase in "experiments per GPU," which in turn shortens the cycle from hypothesis to validation. For data science teams, experiment throughput directly determines the speed of model iteration. Deel's case shows that the resource management capabilities of post-training platforms are becoming a point of differentiated competition—not just "being able to run," but "being able to run many at once."

The reason GPU resource reuse and multi-experiment parallelism are feasible lies in the fact that modern scheduling systems can partition compute along both temporal and spatial dimensions. Spatially, GPU memory partitioning (such as MIG, MPS) or multi-process sharing allows smaller-scale experiments to share the same card; temporally, schedulers exploit the peaks and troughs of the sampling and update phases in RL training, interleaving tasks from different experiments to reduce GPU idling. In addition, RL experiments often involve substantial CPU-side rollouts, environment interaction, and reward computation, so the GPU is not fully utilized throughout—this provides natural gaps for stacking multiple experiments. The traditional "one experiment monopolizes a set of GPUs" model causes significant waste, whereas platform-based scheduling converts this fragmented idle time into effective compute, thereby increasing "experiments per GPU."

Implications for RL Engineering Practice

Although Deel's sharing was brief, it points to a clear trend: the infrastructure for RL training is moving from a "handicraft workshop" to "platformization." In the past, teams had to maintain their own training frameworks, scheduling systems, and monitoring tools; now, post-training platforms productize these capabilities, allowing data science teams to focus their energy on experiment design and results analysis.

This shift is especially meaningful for small and medium-sized teams. Without a large ML infra team, they can still leverage platforms to achieve experiment throughput close to that of major companies. Of course, platformization also brings new considerations: the degree of dependence on the platform's abstraction layer, the portability of experiments, and cost transparency are all factors teams need to weigh when choosing.

The Value of a Fast Iteration Loop

One sentence in the original text captures the core logic: "more experiments per gpu means a much faster iteration loop." In the RL field, experiment iteration speed is often more important than the scale of a single experiment, because RL has low sample efficiency and is sensitive to hyperparameters, requiring extensive trial and error. Teams that can run experiments at lower cost and with higher parallelism are more likely to quickly find effective training recipes.

Deel's case provides a reference benchmark: setup completed in a few hours, and five-fold parallel experiment capacity. For teams evaluating post-training platforms, this is an efficiency signal worth paying attention to.

(This article is based on public information shared by the Deel data science team on social media and does not include independent verification by the platform provider or third parties.)

One-line Summary

Deel team ran RL experiments within hours using the Together platform, running five experiments in parallel on a single GPU capacity.

Paragraph Summary

The article presents a case study of the data science team at fintech company Deel using Together's post-training platform for reinforcement learning experiments. The team completed the entire process from environment setup to real RL experiments within hours, far faster than the traditional setup cycle of days or even weeks; they then ran up to five experiments simultaneously on the same GPU capacity, significantly improving experiment throughput and effective GPU utilization. The article notes that this reflects RL training infrastructure shifting from a "handicraft workshop" model—where teams build their own frameworks, scheduling, and monitoring—to a "platformized" model in which platforms productize and encapsulate these capabilities, enabling small and medium-sized teams to achieve experiment iteration speeds close to those of major companies. It also cautions about new considerations brought by platformization, such as dependence, portability, and cost transparency, and emphasizes that in the RL field—where sample efficiency is low and hyperparameters are sensitive—a fast iteration loop is more valuable than the scale of a single experiment.