Team
Small team.
Dense output.
MingLLM is founder-led from Palo Alto, with a disproportionate opinion about what personal intelligence should look like. Today that means one person carrying model training, product, and engineering — and shipping only the parts we can stand behind ourselves. We're building out the early team now.
Builds MingLLM end to end — model training, product design, and the engineering that makes a small on-device model feel like frontier on your own machine. Voice (Jarvis), browser (Tensor), the Loom coding agent, and Rocky all run through one hand.
We're hiring.
Research
Model training engineer
MoE routing, expert iteration, eval-driven loops. You've trained or tuned a small model end-to-end and have opinions about why frontier-scale isn't the answer for personal agency.
Engineering
Client engineer · macOS
Swift + Metal + a steady hand for system-level integration. The Jarvis voice loop, the Tensor side panel, the receipts log — anything the user actually touches.
Engineering
Inference + tooling
MLX kernels, custom ops for the routing layer, and the dev tooling that makes the training loop tight. You like making one thing 4× faster more than you like adding two things.
Generalists who do not see themselves in any of these are still interesting to us — same address (yimingbeckmann@mingllm.com), just say so.