Back to home
MoE architecture
3 articles tagged with this topic
Moonshot AIDeepSeek
2.8T-Parameter Model Crammed Into Gaming PC — Runs, But Too Slow to Use
Engineer runs 711GB, 2.8T-param Kimi K3 on RTX 4070 Ti + 32GB RAM — normal speed 1-2 tok/s. Dual SSD parallelism pushes DeepSeek from 1.14 to 8.08 tok
4d ago2 min read
AlibabaQwen
Alibaba's Qwen Frustrates Local AI Users — They Want a Runnable Version
A Reddit user praises Qwen 3.8 27B quality but can't run it on M1 Max. They want 35B A3B — slower but usable. Local LLMs hit a hardware wall.
6d ago2 min read
AI9StarsG9v3-39A5B
39B Open-Source LLM Runs on a PC — Without the Vendor's Help
AI9Stars' G9v3-39A5B was quantized to GGUF by Reddit user linuxid10t via a llama.cpp fork—open-source LLM tooling still runs on community relay.
Aug 152 min read