Job description

ABOUT THINKING MACHINES The mission of Thinking Machines is to build AI that extends human will and judgment. We are training frontier models with Inkling, developing Tinker to let people make models their own, and crafting interfaces that broaden human-AI communication. We believe the future worth building is human, and we're hiring people who want to build it. ABOUT THE ROLE We’re looking for an infrastructure research engineer to design and build the core systems that enable efficient large-scale model training with a focus on numerics. You will focus on improving the numerical foundations of our distributed training stack, from precision formats and kernel optimizations to communication frameworks that make training trillion-parameter models stable, scalable, and fast. This role is ideal for someone who thrives at the intersection of research and systems engineering: a builder who understands both the math of optimization and the realities of distributed compute. Note: This is an "evergreen role" that we keep open on an on-going basis to express interest. We receive many applications, and there may not always be an immediate role that aligns perfectly with your experience and skills. Still, we encourage you to apply. We continuously review applications and reach out to applicants as new opportunities open. You are welcome to reapply if you get more experience, but please avoid applying more than once every 6 months. You may also find that we put up postings for singular roles for separate, project or team specific needs. In those cases, you're welcome to apply directly in addition to an evergreen role. WHAT YOU’LL DO - Design and optimize distributed training infrastructure for large-scale LLMs, focusing on performance, stability, and reproducibility across multi-GPU and multi-node setups. - Implement and evaluate low-precision numerics (for example, BF16, MXFP8, NVFP4) to improve efficiency without sacrificing model quality. - Develop kernels and communication primitives that use hardware-level support for mixed and low-precision arithmetic. - Collaborate with research teams to co-design model architectures and training recipes that align with emerging numeric formats and stability constraints. - Prototype and benchmark scaling strategies such as data, tensor, and pipeline parallelism that integrate precision-adaptive computation and quantized communication. - Contribute to the design of our internal orchestration and monitoring systems to ensure that thousands of distributed experiments can run efficiently and reproducibly. - Publish and share learnings through internal documentation, open-source libraries, or technical reports that advance the field of scalable AI infrastructure. SKILLS AND QUALIFICATIONS Minimum qualifications: - Bachelor’s degree or equivalent experience in computer science, electrical engineering, statistics, machine learning, physics, robotics, or similar. - Understanding of deep learning frameworks (e.g., PyTorch, JAX) and their underlying system architectures. - Thrive in a highly collaborative environment involving many, different cross-functional partners and subject matter experts. - A bias for action with a mindset to take initiative to work across different stacks and different teams where you spot the opportunity to make sure something ships. - Strong engineering skills, ability to contribute performant, maintainable code and debug in complex codebases in areas such as floating-point numerics, low-precision arithmetic, and distributed systems. Preferred qualifications — we encourage you to apply if you meet some but not all of these: - Familiarity with distributed frameworks such as PyTorch/XLA, DeepSpeed, Megatron-LM. - Experience implementing FP8, INT8, or block-floating point (MX) formats and understanding their numerical trade-offs. - Prior contributions to open-source deep learning infrastructure such as PyTorch, DeepSpeed, or XLA. - Publications, patents, or projects related to numerical optimization, communication-efficient training, or systems for large models. - Experience training and supporting large-scale AI models. - Track record of improving research productivity through infrastructure design or process improvements. LOGISTICS - Location: This role is based in San Francisco, California. - Compensation: Depending on background, skills and experience, the expected annual salary range for this position is $350,000 - $475,000 USD. - Visa sponsorship: We sponsor visas. While we can't guarantee success for every candidate or role, if you're the right fit, we're committed to working through the visa process together. - Benefits: Thinking Machines offers generous health, dental, and vision benefits, unlimited PTO, paid parental leave, and relocation support as needed.

Requirements

Must-haves

This role is ideal for someone who thrives at the intersection of research and systems engineering: a builder who understands both the math of optimization and the realities of distributed compute.
Design and optimize distributed training infrastructure for large-scale LLMs, focusing on performance, stability, and reproducibility across multi-GPU and multi-node setups.
Implement and evaluate low-precision numerics (for example, BF16, MXFP8, NVFP4) to improve efficiency without sacrificing model quality.
Develop kernels and communication primitives that use hardware-level support for mixed and low-precision arithmetic.
Collaborate with research teams to co-design model architectures and training recipes that align with emerging numeric formats and stability constraints.
Prototype and benchmark scaling strategies such as data, tensor, and pipeline parallelism that integrate precision-adaptive computation and quantized communication.
Contribute to the design of our internal orchestration and monitoring systems to ensure that thousands of distributed experiments can run efficiently and reproducibly.
Publish and share learnings through internal documentation, open-source libraries, or technical reports that advance the field of scalable AI infrastructure.
Bachelor’s degree or equivalent experience in computer science, electrical engineering, statistics, machine learning, physics, robotics, or similar.
Understanding of deep learning frameworks (e.g., PyTorch, JAX) and their underlying system architectures.
Thrive in a highly collaborative environment involving many, different cross-functional partners and subject matter experts.
A bias for action with a mindset to take initiative to work across different stacks and different teams where you spot the opportunity to make sure something ships.
Strong engineering skills, ability to contribute performant, maintainable code and debug in complex codebases in areas such as floating-point numerics, low-precision arithmetic, and distributed systems.
Familiarity with distributed frameworks such as PyTorch/XLA, DeepSpeed, Megatron-LM.
Experience implementing FP8, INT8, or block-floating point (MX) formats and understanding their numerical trade-offs.
Prior contributions to open-source deep learning infrastructure such as PyTorch, DeepSpeed, or XLA.
Publications, patents, or projects related to numerical optimization, communication-efficient training, or systems for large models.
Experience training and supporting large-scale AI models.
Track record of improving research productivity through infrastructure design or process improvements.
Location: This role is based in San Francisco, California.
Visa sponsorship: We sponsor visas. While we can't guarantee success for every candidate or role, if you're the right fit, we're committed to working through the visa process together.

Nice-to-haves

Preferred qualifications — we encourage you to apply if you meet some but not all of these:

Application prep

Use the role’s language and application details while tailoring your materials.

Keywords to mirror

aimachine learning

Platform: ashby
Effort: quick

Benefits

Benefits detected in the listing.

Dental

Vision

Pto

Parental Leave