Raon-1 nano tops consumer-GPU agent tests: 96/96 tasks, 187 tok/s at ~2GB Q4

Kangwook_Lee · x · 2026-11-13

Blogger PaulGug benchmarked Krafton's new Raon-1 nano against Qwen 3.5 2B, Gemma 4 E2B, MiniCPM5 2B and LFM3 1.7B on an RTX 3080 (10GB, 51K context) with identical llama.cpp settings.
Full agent tasks (PA Core, reasoning off, out of 300): Raon-1 nano led at 163, ahead of MiniCPM5 2B (150), LFM3 1.7B (149), Gemma 4 E2B (124) and Qwen 3.5 2B (114).
Individual skills: MiniCPM5 2B leads tool calling (99%) and long context (92%); Raon is lower there (86%/75%) but strong on instruction following (93%) — the best single-skill model wasn't the best end-to-end agent.
Raon-1 nano completed all 96 agent attempts with zero failures, runs 187 tok/s, and is only 2GB at Q4 (7GB VRAM free), making it the pick for multi-step agent work; MiniCPM5 2B wins tool-heavy or long-context use. With reasoning on, most small models exhausted their thinking budget — Gemma E2B was the exception (154).

More from coding & agent