Sabitlenmiş Tweet
With local LLMs, like everywhere, hype X accounts are confusing things with bullshit.
Here are some facts I learned so far:
💎 No, 32 or 36 GB unified RAM is not enough. Ideal minimum is 64 GB.
💎 Model decision always comes down to Gemma 4 31B (or 26B MoE) or Qwen 3.5 27B (or 35B MoE). These are the only multi-use models. Question is, can you make them fast?
💎 Yes, MLX on Apple Silicon is the real deal. Without it, local LLMs are just a stunt.
👀 If you are a beginner wanting to use an LLM in GUI, use LM Studio directly and forget about other tools.
🕸️ Open WebUI works just as well with LM Studio, which is good because it gives internet access abilities to your LLM (making it actually usable).
🤯 No current local LLMs will be able to run advanced skills with OpenClaw or Hermes. At best, they can be the brains of single-use agents, nothing more.
There is a lot of hype bullshit on X. Do not fall for it; it will cost you time and money. If you need agents that get something done, accept the cost for tokens and use cloud frontier models such as Codex or Claude directly. You will not save money by buying a MacBook Pro M5 Max.
English
