On the 6th, Nota unveiled two lightweight versions of MoonShot AI's open-weight AI model 'Kimi K3'. In two versions with expert modules reduced by 25% and 50% respectively, the number of Nvidia B300 GPUs required for operation decreased from the original 8 units to 6 and 4 units respectively, the company announced.

 

 

Kimi K3 adopts a Mixture-of-Experts (MoE) structure that selectively activates only necessary modules based on input values. Because weights and activation values were designed as 4-bit and 8-bit respectively during the training phase, the margin for additional compression using conventional quantization methods was limited.

 

Nota explained that it resolved this challenge with 'Non-uniform Expert Pruning' technology. By analyzing layer-specific characteristics and module importance, the company selectively removes only low-contribution modules while applying different removal ratios for each layer to minimize performance degradation.

 

According to proprietary benchmark results, Nota's 50% lightweight model achieved higher performance retention rates across most benchmarks compared to the REAP (Router-weighted Expert Activation Pruning) method, which removes experts at equal ratios per layer. The high-difficulty inference evaluation metric GPQA-Diamond and instruction-following capability metric IFEval scores were ahead by 2.52 and 2.22 points respectively.

 

 

Nota's previously released lightweight model of Upstage 'Solar Open2' on Hugging Face exceeded 68,000 cumulative downloads within two weeks of launch. The Kimi K3 lightweight model can also be found on the same platform.

 

Taeho Kim, Chief Technology Officer (CTO) of Nota, stated, "Following Solar Open2, we have reduced the required GPUs by up to half while maintaining core performance for Kimi K3 as well," and added, "We will expand the scope of technology application to enable efficient utilization of various frontier-class AI models even within limited computational resources."