el.cine (@EHuanglu)

Kimi k2.5 모델을 로컬에서 실행해 2대 Mac Studio 환경에서 초당 약 24토큰 성능을 기록했다는 보고. 이는 곧 웹 기반 구독 없이도 로컬에서 실용적인 모델 운용이 가능해질 수 있음을 시사한다.

https://x.com/EHuanglu/status/2016611220803969125

#localllm #kimi #macstudio #edgeinference

Replicate joins Cloudflare in a massive bet on the future of developer friendly AI

https://fed.brid.gy/r/https://nerds.xyz/2025/11/replicate-joins-cloudflare/

New benchmark results show Weibo's VibeThinker‑1.5B outperforms DeepSeek‑R1, costs just $7.8K, and matches larger models on GPQA math and code tasks—while running on edge devices. Curious how this shifts inference economics? Dive into the full analysis. #VibeThinker15B #DeepSeekR1 #GPQA #EdgeInference

🔗 https://aidailypost.com/news/weibos-vibethinker-15b-beats-deepseek-r1-usd-78k-ties-larger-models