Nemotron 3.5 Bolt E4B hits 27.7 tok/s on Jetson Orin Nano Super, 1.8x faster than baseline
Any_Frame9721 · reddit · 2026-11-06
The Embedl team got NVIDIA's new Nemotron 3.5 Bolt E4B running at 27.7 tok/s on a $399 Jetson Orin Nano Super (8GB), versus the official 15 tok/s llama.cpp baseline — a 1.8x speedup using their TensorRT Edge LLM variant with int4 quantization. The model runs fully on-device and can code small games; you can try it live on their Jetson device farm or run the TensorRT engine locally.
More from Embodied
- Tesla reportedly doubling Giga Texas output with night shift to ramp Cybercab — mitchdeg · 2026-11-06
- Princeton/DeepMind professor's free 24-lecture robotics course, plus how π0, GR00T and Optimus think — Majumdar_Ani · 2026-11-06
- NVIDIA open-sources cuRoboV2-Fast: robot motion planning cut from 100ms to 4.4ms — ankurhandos · 2026-11-06
- China's robotics flywheel: scale, cost, standards and field data — ingliguori · 2026-11-06
- Liquid AI's LFMs roll out in Mercedes-Benz vehicles, running fully on-device — JosephJacks_ · 2026-11-06
- GPT-6.1 Sol halves surgical positioning error after 3 mouse brain surgeries — 141_1337 · 2026-11-06