Research Engineer - Pre-training

Job description

Pluralis Research works on Protocol Learning: training and serving large models in a fully decentralized way on small consumer-grade devices connected via the internet. Despite being dismissed as infeasible, we have made significant advances on this problem, most recently Agora, a permissionless run that pretrained an 8B model from scratch on consumer GPUs spread over the internet, with no single participant ever holding the full weights ([tech report](https://arxiv.org/abs/2607.13332)). While many of the core research problems have been solved, Protocol Learning unlocks a series of new challenges. For the mission in full, read [A Third Path: Protocol Learning](https://pluralis.ai/blog/a-third-path-protocol-learning/). This setting breaks nearly every assumption of datacenter training: communication-efficient training across different parallelism axes, fault tolerance as nodes join and drop mid-run, heterogeneous compute and networks, and robustness to malicious participants. Our published methods include [Subspace Networks](https://arxiv.org/abs/2506.01260), [Factored Gossip DiLoCo](https://arxiv.org/abs/2606.22768), [AsyncMesh](https://arxiv.org/abs/2601.22442), and [Sentinel](https://arxiv.org/abs/2603.03592). As a Research Engineer you'll build the training system that takes Protocol Learning from the 8B run to frontier scale: large models on heterogeneous hardware, in physically different regions, connected by ordinary internet. ## Key Responsibilities - **Distributed pretraining**: Implement and optimize model-parallel training. Data, pipeline, and tensor parallelism for large models on heterogeneous GPUs under low-bandwidth, high-latency links. - **Performance optimization**: Implement techniques that reduce communication overhead while maintaining model convergence in challenging network environments. - **Elasticity and fault tolerance**: Make runs survive node churn. Robust checkpointing, state synchronization, and recovery as participants join and leave. - **Run instrumentation**: Build the monitoring that shows throughput, bottlenecks, and model quality across hundreds of devices. ## What We're Looking For - **Hands-on distributed training (required)**: You've trained models across many devices in PyTorch with FSDP, DeepSpeed, Megatron, or your own implementation. You understand data, tensor, and pipeline parallelism. - **Strong engineering**: Production-quality Python. Concurrency, failure handling, profiling before optimizing. - **Evidence of execution**: Shipped systems, research code, open-source work, or serious personal projects. - **Mission alignment**: You believe Protocol Learning is the viable third path for collective, trustless, and sovereign AI. ## Nice to Have - Hands-on experience training or serving large language models such as Nemotron, Qwen or OLMo. - Experience with P2P networking and NAT traversal. - Experience with post-training and RL. - Experience with inference and serving systems. - Experience at proprietary, open-weight and open-source AI labs ## Compensation & Benefits - **Equity-Heavy Package**: We offer significant ownership for key technical contributors in addition to a high base salary. - **Remote-First Culture**: Flexible work environment with team members distributed globally. - **Visa Sponsorship**: Optional full visa sponsorship and relocation support to either Australia or the US. - **Open Problems**: Training and serving frontier models on hardware you don't control, over networks you don't own, mostly has no published answers yet. You'll write some of the first ones. ## FYI's - We work remotely across the world, with the main teams in Australia and North America. You'll need to be comfortable working across timezones. - Applicants must have professional-level English proficiency (written and spoken). - Recruiters: we aren't looking for agency support at this time. We'll reach out if we need help. *We are backed by* [*Union Square Ventures*](https://www.usv.com/) *and other tier-1 investors, and we are a world-class, deeply technical team of ML researchers. Pluralis is unapologetically ideological. We believe AI, and the world, end up on a better path if we succeed in implementing the protocol for intelligence. If this resonates, please apply.*