Magic builds large-scale code generation models with heavy infrastructure investment: GPU/TPU compute (CUDA, XLA, Triton), cloud orchestration (Kubernetes, Terraform), and low-level systems work (C++, Rust, NVMe optimization). The engineering-dominant hiring mix and active projects around compute kernel porting, synthetic data generation, and evaluation infrastructure signal a company scaling model training and inference at frontier scale rather than shipping finished products.
Tech Stack 25 technologies
Core StackAWS Terraform Pulumi CloudFormation Kubernetes C++ Go Rust CUDA XLA GPU TPU GCP Azure OCI AWS CDK NCCL Triton Mojo Ruff Ray NVMe NFS cgroups SSD
What Magic Is Building
◆Challenges
- Improving evaluation correctness
- Reproducibility of benchmarks
- Dataset quality measurement
- Navigating unknown product development
- High-throughput data movement
- Strict availability requirements
- Sustained pressure on compute
- Optimizing memory utilization for long-context windows
- Data movement optimization for long-context windows
- Sustained memory pressure
▲Active Projects
- Evaluate porting magic’s compute kernels to alternative hardware options
- Internal evals platform
- Improve experiment reproducibility and orchestration workflows
- Dataset quality measurement system
- Post-training dataset acquisition
- Synthetic dataset generation
- Data pipeline infrastructure
- Command-line tool for ltm code pairing
- Cloud-based development environment for ltm
- Large-scale evaluation infrastructure
Hiring Activity
Steady10 roles · 2 in 30d
Department
Seniority
Company intelligence