colibri v1.11.0 runs DeepSeek V4.1 Flash — 552B params, 510 GB on disk — on a CPU box by streaming experts off an SSD

GitHub·high signalXBlueskyLinkedInCopy link

Released 2026-09-13, colibri's ninth engine reads the vendor checkpoint natively with no conversion: fp8 dense with 32x32 ue8m0 tiles and fp4 experts laid out byte-identically to the mxfp4 its Kimi K3 engine already reads. During development a cold turn went from 78.7 s to 25.1 s (0.305 to 0.957 tok/s) through batched expert reads (V41_READ_DEPTH, default 8 is the measured knee), reading attention matrices once per block of positions instead of per token, and an expert-major MoE; a five-turn chat runs 1.14 to 1.58 tok/s. Two approaches to hiding expert reads behind matmuls were built, measured worse, and removed, with the numbers written up in docs/deepseek-v41.md so the next attempt starts from them.