Voxel pagoda generated fully local on a GPU-less 2020 ThinkPad with a 35B MoE model
dai_app · reddit · 2026-11-12
A Redditor ran the viral voxel pagoda prompt fully local on a 2020 ThinkPad (Ryzen 7 PRO 5850U, 16GB RAM, no GPU) using Qwen3.6-35B-A3B Q4KM — a 21GB model larger than the RAM — with BigMoeOnEdge v0.42's built-in agent. Output: 988 lines of HTML + Three.js, 11.8k tokens at 2.3 tok/s, roughly 1.5 hours.
It works because of MoE: each token activates only a few experts; dense weights and a 2GB expert cache stay in RAM while the rest streams from the SSD, on unmodified llama.cpp. Bottlenecks: half the time is SSD waits, and every agent turn re-reads the whole conversation (30 min on this laptop). On a 16GB Mac the same test took 20 minutes (20 tok/s decode, 100–1000+ tok/s prefill).
More from Infra
- SpaceX's Southaven data centers net local schools and county $63M a year — 0xAllen_ · 2026-11-12
- Quantum infrastructure firm Q-CTRL announces Series C led by Salesforce Ventures — MJBiercuk · 2026-11-12
- Nebius Q3: $12B in customer commitments, revenue up 68% to $981M as GPU prices climb 22% — demian_ai · 2026-11-12
- 1T-Parameter MoE trained on EU public supercomputer with PyTorch, per-GPU throughput up 14x — PyTorch · 2026-11-12
- llama.cpp crosses 134k GitHub stars as the default local LLM inference stack — ggml-org · 2026-11-12
- TensorFold: a Zig LLM inference engine spanning Metal, CUDA and Vulkan — ashhart · 2026-11-12