Reka details Rho-1 training: from scratch on 320 H100s in ~3 months, a fraction of frontier compute
RekaAILabs · x · 2026-10-05
Reka shared training details for Rho-1: the 19B omni model was trained from scratch on 320 H100s over roughly 3 months — a tiny fraction of the compute behind today's frontier models.
The team says Rho-1 closes the loop for physical AI: the same weights that predict what the camera sees next also dictate joint actuation, with no external planner querying a separate world model.
More from Infra
- Hugging Face ships Datasets 5.1 with Vortex, Harbor RL and five bio formats — lhoestq · 2026-10-05
- boat launches: $20/mo full-VM sandboxes for AI agents with 2K concurrent instances — Scobleizer · 2026-10-05
- Nvidia guides to 70% revenue growth after four blowout quarters, bears stuck in the weeds — firstadopter · 2026-10-05
- Qwen3.8-Flash-Next (125B) runs at 59 tok/s on a single Strix Halo mini PC, engine open-sourced — Yaniss916 · 2026-10-05
- Red Hat AI ships NVFP4 quantized Qwen3.8-Flash-Next: MoE experts in FP4, vLLM-ready — huggingface · 2026-10-05
- Why Your RAM Is So Expensive: A Viral Thread Points to the AI Memory Squeeze — TheMoonMidas · 2026-10-05