kimi-k3-in-c is a from-scratch inference engine for Moonshot's 2.78-trillion-parameter Kimi K3, written in portable C99 with no BLAS, no framework, and no GPU. The headline measurement is that the model generates on a machine with 8.24 GB of peak resident memory, from a checkpoint that occupies 1.56 TB on disk. The entire engine is about 176 KB of code.

The design rests on deciding where each byte lives rather than on shrinking the model. K3 is a mixture-of-experts model, and the ~1.45 TB of routed experts are never made resident: they are read from storage and multiplied straight out of their packed 4-bit MXFP4 form, with 16 of 896 experts selected per token. The dense trunk — 93 layers, packed so each is one read — stays in memory to whatever depth the operator chooses, which turns the memory floor into a dial rather than a hard requirement. An LRU cache holds recently used experts, sized from a recorded routing trace instead of guesswork.

The project is explicit that more memory buys speed, not different answers. On the author's 124-core test machine with a fast NVMe drive, a token takes about 26.5 s at 8 GB, 24.2 s at 32 GB, 19.8 s at 64 GB, and 5.6 s at 128 GB+, where the model finally fits in RAM and the disk wait disappears. Output is reported as byte-identical across every memory budget, which is the invariant the validation suite is built around: a tiny oracle model first, then gate checks against the full checkpoint, then sustained generation.

Two attention reductions carry the rest of the load. KDA gives an attention path whose memory does not grow with sequence length, and MLA compresses ninety-six heads into a single latent. Release v1.0.0 (2026-08-07) reports roughly 8x less arithmetic per token than the initial version, a 3.9x faster follow-up turn in a chat, and about half the cost on long prompts.

The practical caveats are stated up front by the project itself. The gate is storage — you need the 1.56 TB checkpoint locally — the supported platform is Linux x86-64, and the weights are the base model with no chat template, so the demo outputs are continuations rather than replies. At tens of seconds per token this is a demonstration of where the memory floor actually is, not a serving stack. Licensed Apache-2.0.