A model you can hold.

The number on a model's box (31B, 397B, a trillion) isn't what shapes each word. What shapes each word is the parameters actually engaged on that token. A Mixture-of-Experts model routes each token to a slice of itself: a 397-billion MoE might fire only 17 billion parameters per step. A dense model has no slice: every parameter fires, every token.

That's why active capacity, not headline size, is the real currency, and dense is the way to spend all of it. When Qwen shipped a 27B dense model in April 2026, it beat their own 397B MoE flagship on hard coding benchmarks while fitting in about 17 GB, because the dense model brought all 27B to bear and the giant one fired only 17B active. We took the same lesson: Gemma-4-31B, fully dense, all 31 billion on every token. And one coherent computation is also the honest one to read: no routing decision fragmenting the thought, which is exactly what a model built to be interpreted should be.

We didn't want to build on sand. Gemma-4 is open-weight (run it offline, inspect it, keep it) and it's Google-backed, so it isn't a weekend project that vanishes; it's supported, documented, and here to stay. It also arrives as a family: from tiny to large, one architecture, one set of tools. We can move up and down that line, a small sibling for a draft model or a larger one when the work demands it, without rebuilding a thing.

And owning the weights is what makes the rest of the program possible. You can only verify a computation, or read a reasoning process, on a model you actually hold, never on an endpoint someone else controls, meters, and can swap out from under you.

A 31-billion-parameter model in full precision is a server. Quantized to NVFP4 (four bits a weight, at a quality that holds), it's about 18 GB: one GPU, a workstation under your desk. Support for AMD is coming soon on consumer hardware.

NVFP4 isn't just compression. It's the key that turns "needs a cluster" into "runs on your hardware." It's why our model fits on a single card today, and it's the same door that lets one big-memory box hold models that used to need a datacenter tomorrow.

There are two ways to make a model know more: bake it into the weights, or give it a memory. Most teams reach for training when they should reach for memory, so we split the two.

Reasoning is a capability. We shape it with training: the reasoning-trace work that, as the research shows, actually moves the topology of how the model thinks. Knowledge is different. Facts, context, your documents: those belong in memory, retrieved when needed, not fine-tuned in again and again until the model forgets what it was. Train the reasoning; remember the rest.

Most people fine-tune with someone else's framework and hope. We built the whole training stack ourselves, in Mojo, on the same kernel we run inference on, starting from a nanochat-sized clone and building up to a full pipeline. Training and serving share one deterministic, readable substrate. There's no black box in the loop.

That end-to-end control is the groundwork for the real goal: a model that improves itself under a constitution. Not reward-hacking in the dark, but a lawful loop where the model proposes, a hierarchy of judges weighs it against principles it can't violate, and every gain is measurable: you can watch the topology of its thinking move. Constitutional AI, but on a kernel where every step is deterministic and every improvement is something you can read. Nomos all the way down: even self-improvement stays under the law.