Moonshot AI's Kimi K3 is a 2.78-trillion-parameter mixture-of-experts model with a 1.56 TB checkpoint — the kind of model that normally lives on a rack of GPUs. Fareed Khan's kimi-k3-in-c runs it on a single CPU. The engine is written from scratch in portable C99, uses no GPU, no BLAS library and no ML framework, and the whole engine is a 176 KB binary. It is Apache 2.0 licensed and was published on 1 August 2026, with v1.0.0 following on 7 August and v1.1.0 on 2 October 2026.
The trick is streaming. Kimi K3 has 896 experts per layer, but only 16 of them fire for any given token — so the engine streams weights from disk and loads only the experts that are actually needed. Peak RAM on the laptop preset is 8.24 GB, which is within reach of an ordinary machine; the price is disk space (about 1.7 TB of available disk is needed for the checkpoint) and speed. The project's v1.1.0-era README reports roughly 26.5 seconds per token at 8 GB of RAM and about 5.6 seconds per token at 128 GB or more — slow by any practical standard, but the output is byte-identical to the PyTorch reference at every memory budget, which is the point: the engine proves that correctness does not depend on the hardware budget.
Key facts
| Item | Details |
|---|---|
| Repository | FareedKhan-dev/kimi-k3-in-c (GitHub) |
| Creator | Fareed Khan |
| What it is | From-scratch portable C99 inference engine for Kimi K3 |
| Model | Moonshot AI's Kimi K3 — 2.78T-parameter MoE, 1.56 TB checkpoint |
| Hardware | Single CPU; no GPU, no BLAS, no ML framework |
| How it works | Streams weights from disk; only 16 of 896 experts fire per layer |
| Engine size | 176 KB binary |
| Peak RAM | 8.24 GB on laptop preset |
| Disk needed | ~1.7 TB of available disk space |
| Output | Byte-identical to PyTorch reference at every memory budget |
| Speed (v1.1.0-era README, 2 Oct 2026) | ~26.5 s/token at 8 GB; ~5.6 s/token at 128 GB+ |
| Licence | Apache 2.0 |
| Published | 1 August 2026; v1.0.0 (7 Aug), v1.1.0 (2 Oct 2026) |
| GitHub stars | ~7,000 (October 2026) |
What this means — and what it does not mean
This is a teaching artifact and a correctness proof, not a daily driver. Nobody should run Kimi K3 this way to get work done: at 26.5 seconds per token, a single short paragraph takes the better part of an hour. What the project proves is more interesting than what it performs. It shows that a frontier-scale mixture-of-experts model is, at its core, a sequence of arithmetic operations that can be expressed in a few hundred kilobytes of plain C — and that the same model, run with different amounts of memory, produces byte-identical output. For students of systems programming and machine learning, that is a rare, readable window into how these models actually work under the hood.
Two status notes keep the picture current. There have been no commits to the main branch since 2 October 2026, so v1.1.0 is the latest released state; and 26 pull requests are open, most of them filed between 4 and 6 November 2026, which suggests the community is still engaging with the code even while the main branch is quiet. Star counts measure interest, not production use — and for a project like this, the interest is from people who want to learn, not from people who want to deploy.
A note on verification
The speed figures in this post are the newer v1.1.0-era README figures, dated 2 October 2026. An older pre-v1.0 figure (32.69 seconds per token at 8 GB) is superseded and is not used as current. The star count (~7,000) is an October 2026 snapshot and will have moved since. No claim here has been checked against the live repository beyond the 21 November 2026 snapshot.
Source: FareedKhan-dev/kimi-k3-in-c repository (v1.1.0-era README), via verified research. Last checked: 21 November 2026.
FAQs
What is kimi-k3-in-c?
It is a from-scratch portable C99 inference engine by Fareed Khan (Apache 2.0) that runs Moonshot AI's Kimi K3 — a 2.78-trillion-parameter mixture-of-experts model with a 1.56 TB checkpoint — on a single CPU, with no GPU, no BLAS library and no ML framework. The whole engine is a 176 KB binary.
How can one CPU run such a huge model?
By streaming weights from disk. Kimi K3 has 896 experts per layer but only 16 fire per token, so the engine loads only the experts actually needed. Peak RAM on the laptop preset is 8.24 GB, but about 1.7 TB of available disk space is needed for the checkpoint.
How fast is it?
Slow: the v1.1.0-era README (2 October 2026) reports roughly 26.5 seconds per token at 8 GB of RAM and about 5.6 seconds per token at 128 GB or more. The output is byte-identical to the PyTorch reference at every memory budget.
Should I use it as my daily inference engine?
No. It is a teaching artifact and a correctness proof, not a daily driver. At 26.5 seconds per token, a short paragraph takes the better part of an hour. Its value is in showing, in readable plain C, how a frontier-scale model actually works.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!