Founding ML Researcher (Post Training)

Zibra Labs is building the post-training and inference infrastructure that currently only exists at the frontier labs. Our vision is to enable everyone to post-train and serve open weight models at a fraction of the cost with higher capability.

The research side of that vision owns the recipes: how an open weight model gets post-trained cheaply and served cheaply without giving up capability. You will sit next to the team building the runtime, close enough that a recipe needing a new primitive can get one.

What You Will Work On

• Post-training recipes for open weight models that hold up outside the benchmark they were tuned against.
• Reinforcement learning at the algorithm level: reward design, rollout efficiency, and keeping long-horizon runs stable.
• Making a served model cheaper without making it worse, through quantization, distillation, sparsity, and speculative decoding.
• Evaluation that survives contact with reality, catching the regressions a leaderboard number hides.
• Shaping the product direction: what we post-train, what we serve, and what ships next are research calls as much as engineering ones.

About You

• You have trained models at a scale where the infrastructure fought back, and you can tell which failures were the recipe and which were the machine.
• You read the literature closely enough to know which results reproduce.
• You have taste in experiment design: you can pick the smallest run that answers the question.
• You write code other people can run. Research is not an excuse.
• You're high agency: you can take a capability goal, turn it into experiments, and ship the result.
• You care about the quality of your work.

Jargon That Might Be Useful

None of this is a checklist. It is a map of the territory, so you can tell whether it is the territory you want to be in.

• Post-training. SFT, DPO, GRPO, PPO
• RL. veRL, SkyRL
• Training. PyTorch, JAX, FSDP, DeepSpeed
• Inference. SGLang, vLLM, TensorRT
• Efficiency. FP8 and INT4 quantization, distillation, speculative decoding, LoRA
• Evaluation. lm-evaluation-harness, task-specific harnesses

Apply

Send a note about what you have built and what you want to work on, with anything that shows the work: a paper, a repo, a training run you debugged to the bottom.

Our address: careers (at) zibralabs (dot) ai