Surrogate Inference Engine Accused of License-Washing Open-Source Project NInfer

teortaxesTex · x · 2026-11-03

A forensics thread alleges the Surrogate serving engine—famously benchmarked at 7.0x llama.cpp—was built on someone else's open-source work.

Surrogate itself is real: 2,700 commits since January 2, 670k lines of C++/CUDA plus 180k Python, overwhelmingly authored by Claude models (Opus 4.7–5.5), with human contributors accounting for only 45 commits.
The engine was imported wholesale on August 23 from NInfer, an Apache-2.0 project by Shanghai-based developer Neroud; every ninfer identifier was renamed to sinfer, a crediting NOTICE was added and then deleted two days later, and the README no longer mentions NInfer.
30% of engine code lines still match the August NInfer snapshot, with 266 files nearly unchanged—though genuinely new work exists, including Qwen3.8-Flash-Next support six weeks before NInfer's own.
Whether the NOTICE deletion was careless agent cleanup or deliberate remains unclear.

More from coding & agent