Research Engineer - Inference

Prime Intellect is building the open superintelligence stack: the infrastructure frontier AI labs build internally, made available to every ambitious AI team.

Our platform, Lab, unifies compute, environments, evaluations, secure sandboxes, high-performance training, and deployment into one full-stack system for post-training at frontier scale - from SFT and RL to tool use, agent workflows, and continuously improving production models. We are building open frontier AI: open-source models trained end to end for long-horizon tasks like autonomous research, and the full-stack platform our own research team uses to build them. The next generation of AI companies, enterprises, and research teams do not just need more GPUs. They need the ability to turn their own workflows, tools, data, and feedback loops into superintelligence they own.

Prime Intellect has raised $150M in total funding from Founders Fund, Radical Ventures, NVIDIA, and exceptional AI, infrastructure, and enterprise operators, including Andrej Karpathy, Dwarkesh Patel, and leaders and founders from Ramp, Perplexity, Harvey, Mercor, Zapier, Datadog, Cognition, OpenAI, Thinking Machines, Together AI, SemiAnalysis, LangChain, Browserbase, Cloudflare, Sierra, Databricks, Airbnb, OpenRouter, Standard Intelligence, Fleet, Core Auto, and more. We are looking for people who want to build at the intersection of frontier research, real infrastructure, and go-to-market for a category that does not fully exist yet.

Build the inference systems that power Prime Intellect’s RL research, scaling toward 1–3T parameter models on NVL72 clusters. Your work will improve rollout throughput and make frequent model updates practical.

• →

  Extend prime-rl’s vLLM integration to bring new model architectures into RL training, with small-scale validation before larger runs.

• →

  Improve rollout throughput for long-horizon agents, tuning concurrency and KV-cache reuse as context lengths grow.

• →

  Optimize multi-node MoE inference, including expert parallelism and the tradeoffs of separating prefill from decode.

• →

  Reduce trainer-to-inference weight-update overhead through sharding-aware transfer and communication overlap, while keeping rollout workers on the intended policy version.

• →

  Build on FP8 and evaluate lower precision. Use trainer/sampler mismatch checks to catch numerical regressions and measure the effect on RL training.

• →

  Work with model researchers on architecture choices that improve inference efficiency, balancing model quality against execution cost on the target hardware.

• Experience building or optimizing GPU inference systems, with an understanding of inference-engine internals.

• Strong Python skills and experience with PyTorch or a comparable ML framework.

• Ability to profile end-to-end workloads and turn measurements into verified performance improvements.

• Experience debugging numerical or integration issues in ML systems.

• Contributions to vLLM or another inference engine.

• CUDA or Triton kernel work, particularly for low-precision inference.