About Kog
Kog builds a co-designed inference stack for real-time AI agents on standard NVIDIA and AMD datacenter GPUs.
We develop model architectures and inference systems together. Our work includes Laneformer 2B and Delayed Tensor Parallelism (DTP), a Transformer architecture that overlaps communication with useful computation and weight streaming.
Our inference stack generates 3,500 tokens/s per request on 8 AMD MI300X GPUs and 2,100 tokens/s per request on 8 NVIDIA H200 GPUs, in FP16 at batch size 1, without quantization or speculative decoding.
Our next challenge is to accelerate how new AI models are adapted and optimized for different GPU architectures.
On the research side, we explore architectural innovations, model morphing, and approaches that make models more efficient at inference time. We also plan to build internal research agents to help generate ideas, run experiments, and evaluate architectural changes faster.
We're a team of 10 people, including three PhD holders and one PhD candidate.
Try our technology at playground.kog.ai and explore our research on the Kog Labs blog.
What you will work on
You will research and develop model architectures designed around inference constraints, including:
New approaches to attention, routing, residual structures, and parallelism.
Extensions to Laneformer 2B and DTP.
Architecture choices for large MoE models, including communication and execution at inference time.
Model morphing, modifying the structure of trained models while reusing learned weights.
Experiments that connect architectural hypotheses to measurable changes in model quality and inference performance.
Internal research agents to accelerate architectural exploration and experimentation.
You will work on open research questions, implement your ideas, evaluate their results, and contribute to publications and systems built at Kog.
What we look for
We look for researchers who have personally designed or materially changed LLM mechanisms, not only trained or fine-tuned existing models.
A research doctoral background is required. A completed PhD is preferred, but the thesis does not need to be in LLMs.
Strong understanding of Transformer architectures and their underlying mechanisms.
Demonstrated ability to formulate hypotheses, implement experiments, and interpret results.
Experience owning meaningful research decisions and driving work independently.
Strong implementation skills and clear technical reasoning.
Experience with model morphing is an advantage, not a requirement. Previous experience building AI agents is not required, but you should be interested in contributing to this direction.
We review existing research work during the process, including papers, theses, code, or technical projects.
What we offer
Direct impact on Kog's model architecture and inference technology.
The opportunity to take research ideas from hypothesis to implementation and real inference measurements.
Autonomy to explore new architectural directions and publish results.
Close collaboration with engineers working across model architecture, inference, and GPU execution.
Location: Paris, France. Candidates based elsewhere must be willing to relocate to Paris.