Notta Shrinks Kimi K3.1 Pro So It Runs on NVIDIA B200 GPUs

At a glance

Notta trimmed Moonshot AI's large Kimi K3.1 Pro model so it can run on eight NVIDIA B200 GPUs instead of the B300s it originally needed. It removed rarely used parts and used the freed memory for longer context and faster answers.

Key points

1. 1

   Notta pruned rarely used expert modules by up to 37%.

2. 2

   The company says the slimmed model keeps over 99% of the original score on a hard science reasoning test.

3. 3

   The freed memory allows up to 1 million tokens of context.

4. 4

   By Notta's own estimate, purchase costs fall by up to about 20% and cloud rental by up to about 22%.

Why it matters

Companies without access to the newest GPUs may be able to run large AI models at a lower cost.

Also reported by

• e4ds news노타, 키미 K3.1 Pro 최적화로 GPU 구매비용 최대 20% 절감
• www.ebn.co.kr노타, 초거대 AI '키미 K3.1 Pro' 최적화…GPU 구매비용 최대 20% 절감
• 이머니뉴스노타, 초거대 AI 모델 최적화로 GPU 구매비용 최대 약 20% 절감
• 머니투데이노타, 中 문샷AI 모델 최적화…GPU 구매비용 최대 20% 절감 - 머니투데이
• 헬로티노타, 키미 K3.1 Pro 경량화…클라우드 GPU 임대료 최대 22% 절감