codegirl007

BEST Models for 96gb of VRAM / RAM I'm specifically focused on models with good browser use capabilities/good coding/decent vision. I recommend either: - sitegeist: allows the model to do anything codex browser use can do in the background github.com/0xsero/sitegeist - cua github.com/trycua/cua (computer/browser use) -------------- Qwen-3.8-next-flash: - score 40 on AA - Gemini-3.8-flash levels / GPT-6-Luna max level - very fast - good at browser/computer use - need an extra 50gb of nvme/ssd It's a good model, hosting it is a slight bit more weird but again any free model on Opencode can get it running quick. I recommend: huggingface.co/nvidia/Qwen3.… --------------- Qwen-3.8-27b: - Score: 34 on AA - Opus-4.6 perf - speculative decoding on - medium speed - overthinks Very smart, capable of doing pretty much all computer use tasks, outputs good UIs/assets, and generally performs well. Here's a very optimised version, otherwise any 8bit from Huggingface will do. github.com/kacper-daftcode/q… Opencode free/claude/codex can setup the right one with all the details. huggingface.co/turboderp/Qwe… ------------- Ornith-1.5-122B: - 81.2% on terminal-bench-2.1 matching ds4.1 - also pretty fast Post train of Qwen, pushes it's performance a fair bit huggingface.co/Mia-AiLab/Orn… <-- if you give this to codex/claude/kimi whatever they'll get it running fast ------------------ Nanbeidge: - Dense - Very smart - Very slow Worth looking into this more, but i'd recommend the qwens huggingface.co/Nanbeige/Nanb… ============ If you can get one of these models running, please consider asking it to push it's recipe up to github.com/sybil-solutions/l…