Working Student - LLM Quantization & Inference Performance

Role summary

How much numerical precision does an LLM need, and where in the model architecture does it matter most? When does quantization improve runtime, and what does it cost in model quality?

We're looking for a motivated working student to help turn these questions into experiments and experiments into maintainable code. You'll explore optimization methods and build reusable evaluation tooling. You'll work closely with our quantization, compiler, and runtime engineers, and your findings will directly influence how we optimize models for Tensordyne hardware.

What you will do

• Model quality evaluation: Measure how low-precision computation affects performance on coding and reasoning benchmarks. Diagnose where quantization has the greatest impact.
• Quantization research: Implement and compare post-training quantization methods to improve inference efficiency while preserving task accuracy.
• Build reproducible tooling: Automate evaluations, track configurations and results, and make experiments accessible to engineers and AI agents.
• Measure runtime performance: Characterize latency, throughput, and memory usage to understand bottlenecks under different deployment scenarios.

A first project could involve exploring the quality–performance Pareto frontier of a Qwen or LlaMa model. You'll experiment with 16-bit, 8-bit, and 4-bit quantization configurations to find the best runtime-quality tradeoff.

Core qualifications

• Current enrollment in Computer Science, Mathematics, Electrical Engineering, Machine Learning, or a related master's degree program.
• Strong Python and PyTorch skills, including the ability to understand and modify model implementations.
• Effective use of AI coding tools and critical review of their output.
• An experimental mindset: you can formulate hypotheses, design comparisons, and draw careful conclusions from noisy results.
• Clear communication and the ability to present experimental findings and technical tradeoffs to the team.

Strong pluses: experience with quantization, LLM evaluation, performance profiling, or inference optimization.

This role is for you if you thrive on responsibility, learn quickly, and are naturally proactive .

What we offer

Join a deep-tech team building the next generation of AI inference systems. Put your university knowledge into practice with leading open-source LLMs, deepen your understanding of quantization and inference optimization, and build skills in experimental design, software engineering, and AI coding tools. You'll own a focused project, receive close mentorship from experienced engineers, and have room to explore ideas that shape our optimization methods.

Tensordyne's culture was built on the following values

• Put people first. We only succeed when our people succeed.
• Ethics and integrity always. Be open, honest, and respectful of everyone.
• Think Big. Be ambitious and have audacious goals.
• Aim for excellence. Quality and excellence count in everything we do.
• Own it and get it done. Results matter!
• Make each person better together than they would be as an individual.
• Embrace each other's differences, and embrace that there will be differences.

Tensordyne is an equal opportunity employer. We believe diverse teams are better equipped to solve complex problems and build exceptional technology. All qualified applicants will receive consideration for employment without regard to age, color, gender identity or expression, marital status, national origin, disability, protected veteran status, race, religion, pregnancy, sexual orientation, or any other characteristic protected by applicable laws, regulations, and ordinances.

A note to recruitment agencies: Please do not contact Tensordyne employees or leaders regarding this role. We do not accept unsolicited agency resumes and are not responsible for fees associated with unsolicited submissions.

Find Jobs in Germany on Arbeitnow

We show an excerpt and a link, and the full text when the source allows it. Brands and logos shown belong to their owners.

Source: Statistisches Bundesamt (Destatis), table; dl-de/by-2-0; StarIn's calculation