colibri v1.11.0 runs a 552B model off an SSD on a CPU box, and published the two approaches that failed.

The v1.11.0 release reads DeepSeek V4.1 Flash's vendor checkpoint natively with no conversion, 510 GB on disk, fp8 dense with 32x32 ue8m0 tiles and fp4 experts byte-identical to the mxfp4 its Kimi K3 engine already reads. A cold turn went from 78.7s to 25.1s through batched expert reads (default depth 8 is the measured knee), reading attention matrices once per block of positions, and an expert-major MoE. Five-turn chat runs 1.14-1.58 tok/s. Two attempts at hiding expert reads behind matmuls were built, measured worse and removed, with the numbers written into docs/deepseek-v41.md so the next attempt starts from them. More projects should do that.