HAMMAD YOUSUF

FREE RESOURCE · FROM THE "BYTES" REEL

Kimi K3 on 8 GB of RAM: what it really takes

By Hammad Yousuf · 24 October 2026

Kimi K3 is Moonshot AI's 2.78-trillion-parameter open-weight model (Kimi K3 License). kimi-k3-in-c is a portable C99 engine by Fareed Khan (Apache-2.0, no BLAS, no framework, no GPU) that runs it at an 8.24 GB RAM peak by streaming routed experts and dense layers from disk. Current release v1.1.0 (2 Oct 2026).

At a glance · checked 23 Oct 2026

Peak RAM
8.24 GB, with no GPU
Free disk
About 1.7 TB on fast local NVMe
Speed
26.5-32.7 s per token on a 124-core server
Output
Byte-identical from 8 GB to 224 GB of RAM

Verified numbers (checked 23 Oct 2026)

  • 2.78T parameters, 104B active per token, 16 of 896 routed experts per layer per token
  • 1.56 TB checkpoint in 96 shards; about 93% is routed experts
  • 8.24 GB peak RAM, measured on a 124-core server with fast NVMe and RAM capped at 8 GB
  • 26.5-32.7 s per token on that server; about 16 min per token on an 8 GB Jetson Orin Nano with a hard drive
  • Byte-identical output from 8 GB to 224 GB of RAM

Before you start

  • About 1.7 TB free on fast local NVMe (1.56 TB checkpoint + 109 GB packed trunk)
  • x86-64 with AVX2+FMA or ARM64 with NEON; about 10 GB RAM available for the laptop preset (ultra fits about 3 GB, much slower)
  • C compiler with OpenMP (macOS: brew install libomp); Linux, macOS or Windows via MSYS2
  • Download time: about 3.5 h at 1 Gbps, 35 h at 100 Mbps

The safe rule

Check your disk, not your RAM: you need about 1.7 TB free before you download a single byte.

Steps (from the project README)

Run them in order in a terminal, from the folder you cloned into. Each block copies exactly what to paste.

  1. Clone the repository

    step 1 of 7
    git clone https://github.com/FareedKhan-dev/kimi-k3-in-c.git && cd kimi-k3-in-c

  2. Build the engine

    step 2 of 7
    make -j

  3. Run the weightless tests

    No weights or network needed; ends with ALL WEIGHTLESS TESTS PASSED.

    step 3 of 7
    make test

  4. Check your machine first

    Checks toolchain, RAM and disk speed, suggests a preset; needs about 1.7 TB free before any download.

    step 4 of 7
    ./scripts/k3-doctor.sh

  5. Download the model

    Install huggingface_hub with pipx, then run the download script.

    step 5 of 7
    pipx install huggingface_hub
./scripts/download-model.sh ~/k3model

  6. Pack the trunk

    step 6 of 7
    ./scripts/pack-trunk.sh ~/k3model ~/k3trunk

  7. First run (from the README)

    step 7 of 7
    ./bin/k3 ~/k3model --trunk ~/k3trunk --preset laptop --tok ~/k3model --prompt "The capital of France is" --gen 8 --incremental

What it will not do

  • Text only: no vision, no tools, no server, no quality benchmark.
  • The first token takes much longer.

If you just want to use Kimi K3

  • Hosted API at about $3 per 1M input and $15 per 1M output tokens (Moonshot, plus Together, Fireworks and others).
  • Ollama and LM Studio offer it cloud-only.
  • llama.cpp runs it locally from Unsloth GGUFs of 594 GB to 1.56 TB.

Sources checked

The facts on this page were checked against the sources below on 23 October 2026. The engine changes fast, so treat its README as the final word.

  • kimi-k3-in-c: Apache-2.0, C99, no GPU path, 8.24 GB peak RAM, ~1.7 TB free disk, byte-identical output 8-224 GB, 26.5-32.69 s/token on a 124-core NVMe server. Source ↗
  • Kimi K3: 2,779,931,837,184 parameters, 104B active, 16 of 896 routed experts per token, 96 shards, 1.56 TB, not gated. Source ↗
  • Jetson Orin Nano Super 8 GB with external HDD: token ' Paris' in 948.5-950.7 s (mean 949.4 s), peak 2.57 GiB. Source ↗
  • v1.1.0 (2026-10-02) builds natively on Windows alongside Linux and macOS. Source ↗
  • Kimi K3 License: free use, conditions above $20M revenue / 100M MAU. Source ↗
  • Ollama kimi-k3 is cloud-only; llama.cpp added Kimi-K3 text model (PR #26185, merged 2026-08-15). Source ↗

About the video: the reel that sent you here is presented by an AI avatar of Hammad Yousuf.

Keep going