anyone knows why hermes is starting any session with 16K of context on win but only 1K on win ?

1/ 🚀 This shouldn’t fit on a 3090 — but it does. 🧠 Qwen3.6-27B running a full OpenAI-style agent stack on a single RTX 3090 (24GB): 👁 vision · 🔧 tools · ⚡ streaming · 🤔 reasoning · 🧩 MTP n=3 🚫 no feature cuts 📊 50 / 66 TPS @ 218K text 📊 51 / 68 TPS @ 198K vision Built on @Alibaba_Qwen @vllm_project 👇 github.com/noonghunna/club-3…

Using Qwopus3 Q5 and Hermes is not so good, Hermes is quickly running in loop to try to patch but with copaw, i don't have this problem, it looks much better and doesn't go in loop.

wan 2.2 is starting to feel unfair you drop in a single reference video swap the face with one image and the system rebuilds the entire ad shot-for-shot same pacing same structure same emotional beats new avatar brands are using it to clone top-performing videos across every niche without refilming anything i put together a short breakdown showing – how to pick the right source videos – how to avoid uncanny outputs – and how to generate variations that actually convert rt comment “wan” and i’ll send it over (follow for dm)