LLMS

Here's why I shill Droid 24/7 ———- Today Droid single-handedly: 1. Published a REAP of GLM-5 in FP8, there's a reason no one else has done it DSA is still very new: huggingface.co/0xSero/GLM-5-… 2. Found and Fixed an upstream issue with VLLM + DSA + Hopper where GLM-5's kv-cache would need to recompute and spend 20x the time needed, fixed. 3. Created multiple working quantisations on it's own, it tried exl3 and autoround but both failed so resorted to GGUF (autoround 3 bits doesn't work on ampere) huggingface.co/0xSero/GLM-5-… 4. Implemented github.com/0xSero/turboquant within 24 hours of the research paper coming out, tested it across 5090s, 3090s, H100s, and B200s 5. Has been distilling larger models into LoRA to help me test arxiv.org/abs/2505.21835 and it got an 80% prune to be semi-coherent again. 6. Helped my find research papers, clean up slop with the human-writing skill. 7. Got BYOK working with Anthropic, ZAI, Kimi, MiniMax, OpenAI working in Cursor github.com/0xSero/factory-cu… 8. Helped me Implement blog.comfy.org/p/dynamic-vra… 's dynamic loading, only works on a tiny model, but still. ——- I only have to check in on it every 30-45 minutes (I am talking all 8 of my sessions) the thing will run for 16 hours with like 0 prep All this while I am mostly focused on my actual job and tweeting 24/7 Keep in mind each one of these experiments is running on a different server, with different constraints, like I don't understand how I can get such good results here. ——— I love novelty. Which is why I jump around and talking about all these different tools. I have used all of these harnesses and messed around with every feature. I keep coming back to this, and I keep shilling it because I sincerely wish others get to experience this.

→ View original post on X — @nathanlands, 2026-03-27 19:36 UTC