Comparative Evaluation of Small MoE Models in Parallel Coding Agents
A developer benchmarked three recently released small Mixture-of-Experts models—Flint 40B-A2.5B, Ornith 1.5 35B-A3B, and Nemotron 3.5 Lightning—using a parallel coding agent workflow in Three.js. Flint proved to be the most efficient, achieving the shortest wall time, fewest fixes, and lowest API spend despite not having the highest raw speed. This evaluation highlights that raw tokens-per-second performance does not dictate overall efficiency in complex agentic workflows, where edit success rates and error recovery play a larger role. It demonstrates how the choice of local worker models directly dictates the cost and consumption of orchestration APIs like Opus 5.5. The test used Q8_0 GGUF quantizations on an RTX PRO 6000 Blackwell 96 GB card via llama.cpp, executing a complex 1,000 to 1,700-line Three.js task with three parallel workers. Nemotron suffered from a low edit success rate of roughly 17%, repeatedly failing to locate old strings and accumulating stray syntax braces.
## BACKGROUND
Mixture-of-Experts (MoE) is a transformer architecture where only a subset of network parameters (experts) is activated per token, enabling high inference speeds while maintaining large model capacities. GGUF is a binary file format designed for storing quantized large language models to run efficiently on consumer or enterprise hardware via tools like llama.cpp, while Three.js is a cross-browser JavaScript library used to create animated 3D graphics. Atomic Agent is an open-source framework used to orchestrate multi-agent workflows.