About Nunchux AI
Nunchux AI builds infrastructure that makes multimodal generative AI faster and cheaper to serve, and easier to build on. Founded by MIT PhDs Muyang Li, Yujun Lin, Zhekai Zhang, and CMU Professor Jun-Yan Zhu, Nunchux brings together deep research expertise and production systems experience. Built on nearly a decade of research from MIT and CMU, our work includes nunchaku, whose models have surpassed 4 million downloads. We’re backed by top VCs and building for enterprises and millions of developers.
The Role
Nunchux makes visual generative models fast and efficient enough for production. As an ML Engineer on Post-Training & Evaluation, you will build post-training pipelines that improve model efficiency and quality, evaluate the trade-offs, and turn the best recipes into reliable workflows for the models we ship.
What You’ll Do
-
Develop post-training recipes: Establish and validate post-training recipes for image and video generation models.
-
Build data pipelines: Build pipelines to curate and version training and benchmark data for post-training and model evaluation.
-
Build evaluation systems: Build benchmarks and automated judges, and run human preference studies to measure generation quality, fidelity, and efficiency.
-
Benchmark and release models: Benchmark models against relevant baselines, catch quality regressions, and provide clear evidence for release decisions.
-
Bring research into production: Keep up with post-training and evaluation research, then integrate useful methods into the team’s pipelines.
What You Bring
-
Visual generative-model experience: Hands-on experience working with image or video generation models, or other multimodal visual systems. You understand the artifacts, failure modes, and quality trade-offs that matter in generated visual content.
-
Post-training or evaluation depth: Strong experience in either post-training, such as distillation or LoRA, or in visual generative-model evaluation.
-
ML engineering strength: Strong Python and PyTorch skills, with experience building post-training or evaluation code that others can run and maintain.
-
Training systems: Comfortable running and adapting post-training workloads across multiple GPUs using FSDP, DeepSpeed, or similar tools.
-
Experimental judgment: Able to design clean experiments, interpret results, and make practical recommendations from the data.
Bonus Points
-
Experience distilling or fine-tuning large-scale diffusion or video-generation models.
-
Experience building MLLM/VLM-based automated judges or reward models for visual content.
-
Experience with large-scale evaluation datasets, annotation pipelines, or preference data.
-
Experience building agentic visual systems.
-
Contributions to major open-source ML projects.
Why Join
-
Core technical work: Build the systems that let Nunchux make models faster without compromising the quality customers care about.
-
From research to product: Your work will inform model releases, support customization, and run in production rather than stay in a notebook.
-
Technical collaboration: Work closely with a CMU PhD already working on post-training, alongside Nunchux’s research and inference teams.
-
Proven traction: Build on open-source work with more than 4 million model downloads and growing industry partnerships.
-
The team: Work alongside MIT, Berkeley, and CMU researchers and veterans from NVIDIA, AMD, Snowflake, and Adobe.
-
Compensation: $180,000 to $250,000 USD base salary, plus meaningful equity and comprehensive benefits, including health insurance and 401K. Actual compensation will depend on relevant experience, skills, and qualifications.
Location: San Francisco, CA. 4 days in office, 1 day remote.
Start date: As soon as available
Visa: We sponsor H-1B and other work visas for exceptional candidates.
Learn more: nunchux.ai
Apply: Please apply through our Ashby careers page.
Nunchux AI is an equal opportunity employer. We celebrate diversity and are committed to creating an inclusive environment for all employees.
Compensation Range: $180K - $250K