Nota (노타) placed 3rd in the 'Efficient Qwen Competition' at ICML 2026, achieving a 6.978x average in...
The AMW Read
Nota is a known player in model optimization; the competition win is an incremental update but demonstrates significance for inference economics and on-device AI.
Nota (노타) placed 3rd in the 'Efficient Qwen Competition' at ICML 2026, achieving a 6.978x average inference speedup on Qwen3.5-4B using a single NVIDIA A10G GPU. The company combined proprietary quantization, speculative decoding, and sliding-window attention. Two of Nota's papers on MoE-based LLM quantization were also accepted at the ICML AdaptFM workshop.
Why it matters: The result demonstrates that inference optimization—not just model scale—is becoming a key competitive axis as AI moves from training to deployment. Nota's approach, combining multiple techniques to preserve accuracy while dramatically accelerating inference, reflects the industry's growing focus on reducing total cost of ownership (TCO) for LLM serving. This is particularly important for on-device and physical AI applications where compute resources are constrained.
Expert take: This competition win validates the 'context-engineering moat' pattern where companies differentiate through inference-time efficiency rather than larger models. Nota's ability to outperform 40+ global teams on a popular open-weight model signals that specialized optimization expertise can create defensible value, especially as Qwen and similar models gain enterprise adoption. The company is leveraging this achievement to expand global partnerships in on-device AI and physical AI.