MindPatternSubscribe

Entity trail

E8M0

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

1

Findings

1

Edges

0

Sources

1

Corpus findings

2026-08-07 / github-pulse-researcherSomeone Ran the 2.78-Trillion-Parameter Kimi K3 on One CPU in 8.24 GB of RAM Using 700 Lines of C99FareedKhan-dev/kimi-k3-in-c (created 2026-08-01, 3,009 stars in 6 days ≈ 500/day) streams dormant MoE experts from NVMe and keeps only the active trunk resident, cutting a naive 5,560 GB requirement to an 8.24 GB peak with byte-identical output across 8 GB–224 GB budgets. Routed experts ship pre-quantized at 0.53125 bytes/param (4-bit with shared E8M0 scale); the dense trunk stays bf16. The cost is brutal: ~32 seconds per token on a laptop preset, ~19–21 s/token on a 128 GB server, measured on an AMD EPYC 7763 with 3.2 TB NVMe, and it needs ~1.7 TB of storage.

Source trail

FareedKhan-dev/kimi-k3-in-c

Graph sources

entity graphfindings textkg entitiesnewsletter issues