kimi-k3-in-c: a 2.78T model running on one CPU in 8 GB of RAM
kimi-k3-in-c is a from-scratch C99 inference engine that runs Moonshot's 2.78-trillion-parameter Kimi K3 checkpoint on a single CPU with no GPU, no BLAS and no framework, streaming experts from disk so peak memory can be as low as 8.24 GB.
Why it matters
The 176 KB binary produces output byte-identical to the PyTorch reference across memory budgets from 8 GB to 224 GB, turning memory from a hard floor into a speed dial — at the cost of roughly 26-33 seconds per token at the smallest preset.
What you could build with it
A researcher debugging a giant mixture-of-experts model could use this engine's exact, reproducible CPU runs to trace a single token's expert routing on a laptop, without booking GPU cluster time for every experiment.
Does it hold up?
A teaching artifact and correctness proof, not a daily driver: independent writeups confirm the byte-identical runs and the 1.7 TB disk requirement, and the author frames it the same way.
Built with kimi-k3-in-c
- How a 176 KB C binary runs a 2.78T model on one CPU with 8 GB of RAMarticle · Walkthrough of the streaming-expert design, the byte-identical correctness gates, and the honest limitation that 200 tokens takes about 2 hours at the minimum preset.
- Kimi K3 local on 8 GB RAM, no GPU: what it really takesarticle · Practical checklist verified in October: 1.7 TB of free NVMe for the checkpoint, AVX2+FMA or ARM NEON, and the warning that the preset name is not a hardware guarantee.
- Kimi K3 in C on 8 GB RAM: storage, speed and evidencearticle · November explainer separating the three numbers people confuse: the 176 KB binary, the 1.56 TB checkpoint, and the 8.24 GB peak memory.
Learn more
First spotted on github: source.
More AI models
VoiceStudio: open-source, fully-local ElevenLabs alternative for voice cloning, dubbing, and audiobooksdebpalash/VoiceStudio is a desktop app that runs voice cloning, voice design, video dubbing, dictation, transcription…models · JEV 0.75Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73NVIDIA Isaac Lab 3.0 GA + Isaac OS open runtime + Agile ONE SN humanoidMajor Isaac platform expansion at GTC Berlin Oct 21: Isaac Lab 3.0 is generally available as the open-source robot…models · JEV 0.73NVIDIA Nemotron 3.5 Retriever collection: Embed 8B/1B, Rerank 8B, Parse 631MNVIDIA's Oct 21 collection drop for agentic retrieval: Nemotron-3.5-Embed in 8B BF16, 1B BF16 and 1B NVFP4…models · JEV 0.72Kimi K3.1 Pro: Moonshot's 2.8T open-weight flagship officially launchesMoonshot AI officially launched Kimi K3.1 Pro on Oct 23, a mixture-of-experts model with 2.8 trillion total / 104…models · JEV 0.72NVIDIA Cosmos 3.5 world models: now openly availableNVIDIA's Cosmos 3.5 family of open world foundation models for robots, autonomous vehicles and vision agents is now…models · JEV 0.71
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.