KODAITSU · How sovereign AI works

Mixture-of-Experts: how sovereign AI runs on ordinary hardware.

A large model stores a great deal of knowledge. A Mixture-of-Experts model uses only a small slice of it for each word it reads — which is what lets frontier-class intelligence run on a plain server CPU, at the edge, inside your institution. This is the design behind GLM, and the blueprint for KODA's sovereign models.

01

Big models, small machines.

Frontier models hold hundreds of billions of parameters. Loading all of them into fast memory needs a rack of GPUs — expensive, scarce, and tied to a cloud you do not control. For a hospital, a factory or a sovereign deployment, that is the wrong shape.

Mixture-of-Experts changes the economics: keep the knowledge broad, but activate only the fraction relevant to each token.

Store broad. Activate narrow.
02

Dense vs. Mixture-of-Experts.

A dense model runs every parameter for every token. A MoE model has many "experts" but a router picks only a few per token — so most of the model stays idle and cheap.

DENSE — every parameter runs 100% active · full compute · full memory MoE — router picks a few ≈5% active Same knowledge stored · a fraction computed per token
active this tokenstored, idle

GLM-5.2: 744B parameters stored · ~5.4% active per token (verified from the Colibri runtime).

03

Inside a GLM Mixture-of-Experts layer.

For each token, a small router scores the experts and picks the top few. A shared expert always runs — it carries the general competence — while the routed experts add specialisation on top. The results are combined.

token Routersigmoid + bias expert 2 ✓ expert 7 expert 40 ✓ expert 128 …hundreds of small experts, top-k picked Shared expert — always on + next representation

Why this design wins: the router balances expert load with a learnable bias — no auxiliary loss fighting the main objective — so experts stay balanced and quality improves. Attention is compressed too (MLA), shrinking the memory cost of long context by roughly 57×.

Store the knowledge in many small experts. Compute only the few the token needs.
04

The memory trick: keep experts on disk, fetch the ones you'll need.

If only a few experts run per token, they need not all sit in expensive fast memory. A sovereign runtime treats VRAM, RAM and disk as one hierarchy: hot experts stay fast, cold experts live on NVMe, and a prefetcher predicts which experts the next layer will call and pulls them in ahead of time.

FAST · VRAM / RAMhot experts, pinned by usage WARM · RAMrecently used + prefetched COLD · NVMe (disk)the bulk of the model, streamed on demand PILOT predicts next layer's experts Result 744B model in ~25 GB runs on a desktop CPU no GPU cluster no external cloud routing is ~72% predictable one layer ahead
fastwarmdiskprefetch

This is the Colibri runtime pattern — Apache-2.0, single-file C, zero dependencies. KODA studies it as the reference for CPU-native serving.

05

KODA's design: a shared base, specialist adapters, sovereign runtime.

We do not chase a 744B model. We apply the same principles at sovereign scale: one shared base carries general competence; small identity adapters make it Koda, or a clinical or industrial specialist; governed memory holds the facts; and a CPU-native runtime serves it all — beside the equipment, inside the institution.

Governed memory (Mnemosyne) Shared base · 2–3Bgeneral competence Koda / Hiro … Clinical Industrial Legal / Code identity + domain adapters — one loaded at a time CPU-native sovereign runtime — Intel · AMD · Arm · edge appliance

Sovereign

Runs on your own hardware. No dependence on a single GPU vendor or foreign cloud.

Persistent

Each agent keeps its identity and memory across sessions — and we measure that it survives.

Governed

Every recommendation is attributable, over-rideable and recorded. Authority stays human.

Efficient

Store broad, activate narrow: frontier-style capability at a fraction of the running cost.

Store broad. Activate narrow. Serve sovereign.

Mixture-of-Experts is why the same idea that powers a 744-billion-parameter model also lets a governed KODA agent live on a factory-floor appliance or a hospital's own server. Frontier design — brought down to where the work actually happens.

Honest boundary: figures cited (744B / ~5.4% active / ~57× KV / ~72% routing predictability) are verified from the GLM-5.2 architecture and the Colibri runtime. KODA's sovereign models are in the Track-D research and product programme — not a claim of a finished product.