Large Language Model Efficiency Researcher
Active AI foundation-model company offering multimodal models, consumer AI products, and an enterprise/developer API platform.
Funding history
About MiniMax Group Inc.
MiniMax develops proprietary multimodal foundation models spanning language, video, speech, and music, and operates AI-native products including MiniMax Agent, MiniMax Design, MiniMax Audio, Talkie, Hailuo AI, and an Open Platform.
Skills
Candidate Availability
Required and preferred rules are kept separate and reflect the wording in the original posting.
About the Role
You will evaluate and improve large-model efficiency from model algorithms, systems software, and hardware perspectives. You will help design model architectures, build performance and cost models, assess quality-efficiency trade-offs, validate hypotheses through prototypes and experiments, and support production deployment.
Requirements
- Understand large language model structures and computation, including Transformer, Attention, and MoE
- Have experience with large-scale model training or inference systems
- Understand GPU or other AI accelerator architectures and performance optimization
- Understand distributed computing and common parallelization strategies
- Use Python and PyTorch for engineering, profiling, and experimental validation
- Quantitatively balance model quality, training cost, inference latency, throughput, reliability, and implementation complexity
Responsibilities
- Participate in model architecture design from hardware, distributed training, and inference-system perspectives
- Build end-to-end performance and cost evaluation methods for computation, memory, communication, and storage
- Evaluate trade-offs between model efficiency and model quality
- Drive model and hardware co-design by identifying architectural bottlenecks
- Assess the compatibility of model structures with training and inference strategies
- Validate architectural hypotheses through performance models, microbenchmarks, prototypes, and model experiments
- Collaborate with algorithm, kernel, training, and inference engineering teams to deploy solutions at scale