DeepSeek-V2 mixture-of-experts inference on the CPU, in C++20, with no third-party libraries.
A 15.7-billion-parameter model normally needs 32–80 GB of GPU memory. LiteMind runs one on a laptop with no GPU at all.
It works because DeepSeek-V2-Lite is a mixture of experts. Each of its 26 MoE layers holds 64 independent feed-forward experts, and a small router picks 6 of them per token — so roughly a tenth of the expert weights decide any single token. LiteMind memory-maps the 29.3 GiB checkpoint on the SSD and lets that routing decision drive what is paged into RAM. A token pays for six experts, not for the model.
| Parameters | 15.7 B in the checkpoint, 2.45 B active per token (15.6%) |
| Checkpoint | 29.3 GiB, memory-mapped — load time 0.2 s, because nothing is copied |
| Always resident | 2.44 GiB; the other 26.81 GiB streams as the router asks |
| Requires | A 64-bit CPU and an SSD. No GPU, no BLAS, no vcpkg, no Python |
Setup and prerequisites, the architecture one page at a time, the measured
numbers, the command line, and troubleshooting. The site is built from
docs/, which reads just as well here on GitHub.
Ragavan M — full-stack and mobile app developer. B.E. Computer Science and Engineering, Government College of Technology, Coimbatore.
Portfolio · LinkedIn · GitHub · Profile on the docs site
Other work — My GCT, a set of tools for students at the same college:
| My GCT Hub | Question papers, notes and syllabus, on Google Play |
| My GCT | The same material on the web |
| My GCT Store | Where students publish and download each other's apps, sites and tools |
| My GCT Slides | Share slides with a four-digit code — no app, no login |
MIT. See LICENSE. The model weights are
deepseek-ai/DeepSeek-V2-Lite
and carry their own licence.