2.78T-param Kimi K3 runs on a single CPU with 8.24GB RAM in dependency-free C99
FareedKhan-dev · github · 2026-11-04
Trending GitHub project kimi-k3-in-c by FareedKhan-dev (10k+ stars) runs inference for a 2.78-trillion-parameter Kimi K3 on a single CPU using just 8.24 GB of RAM.
- Written in portable C99 with zero dependencies: no BLAS, no framework, no GPU
- Leverages AVX2 SIMD, mxfp4 quantization, mixture-of-experts and linear attention
- A from-scratch systems-engineering effort demonstrating extreme memory optimization for massive MoE models
More from Infra
- Dev muses on a distributed model where every home runs its own tiny data center — rickasaurus · 2026-11-04
- GLM-5.3 goes into production on T-Systems' sovereign Industrial AI Cloud in Munich — jietang · 2026-11-04
- Creative Strategies: 7 research notes on AMD, Apple, Marvell-Lumentum, data center outlook — BenBajarin · 2026-11-04
- Training run killed after gradient spikes to 1.64M and loss hits 12.17 — francoisfleuret · 2026-11-04
- Elections to slow new data center buildouts, boosting incumbent hyperscalers' install base — RihardJarc · 2026-11-04
- Serverless SurrealDB on Cloudflare scans 4.5B rows — 350GB — in one second — mattturck · 2026-11-04