Serious AI.
Built here.

StudioC.ai builds private AI systems that think in context, work in your world, and stay under your control.

Harness · live footageNo concept renders

Harness Deep research
0:00 0:00
Harness Cited answer
0:00 0:00
Harness Analysis
0:00 0:00
Harness Meeting actions
0:00 0:00
Harness Scheduled work
0:00 0:00

What we build

AI that works
the way you do.

Private by design. Practical by default. Built for teams who need answers, not noise.

01 · Quantisation & recovery

Project Anneal

A weight-encoding format, and a recovery-training pass that puts back what the compression cost.

Annealing is heating and slowly cooling a metal so its structure reforms stronger — which is what the recovery pass does to a compressed model.

Two halves. The first is an encoding that stores weights at under two bits each without the collapse that normally causes. The second is a short recovery-training pass against the full-precision model, which buys back accuracy the encoding gave up. Neither needs the original training data.

1.5-bit class
94.6% of full-precision quality retained at 1.71 bits per weight.
1-bit class
89.5% retained at 1.125 bits per weight.
Answer-level check
85.71% of the reference model’s correct answers preserved at ~1.61 effective bits — scored answer-by-answer against the same questions, not by aggregate score.
Recovery training
A short pass against the full-precision model recovers accuracy the encoding cost — 85.71% of reference-correct answers preserved after recovery, up from 84.69% before it, at the same file size. That was proven on a run of roughly 9,000 updates — a 500-update core plus 8,500 teacher-recovery steps. Scaling that to a multi-hour run is the single biggest lever we have, and it is untested only because it needs compute we do not yet own.
Against the wall
Standard binary quantisation collapses at this range — perplexity 1,487,073 at 1.25 bits. Ours holds 104.8 at 1.085 bits, a lower bit rate. Training-free.

02 · Optical architecture

Project Refract

A wave-state model built for optical processors, running today on silicon.

Refraction is light bending as it changes medium — the same physics the optical hardware computes with.

Conventional attention costs more as the conversation grows, which is why long context is expensive and short context forgets. Refract carries a fixed-size wave state instead, so cost stays flat however long the window gets. It is designed for optical processors, where wave operations are the native primitive — the numbers below were measured on silicon, and the gains on optical hardware should be considerably larger.

Quality held
97.4% of the full-size teacher on paired prompt-likelihood, at half the weight precision.
Working memory
5,000,000 tokens processed against a fixed-size core state — the state does not grow with the conversation.
Fits a workstation
Target runtime under 32 GB resident, weights and state included, on a single Apple Silicon machine.
Context floor
262,144 tokens before the memory layer is even engaged.

03 · Agentic runtime

The Harness

No training. Better results.

A harness carries a load without altering what pulls it — the model underneath is never changed.

The Harness never touches the weights. It keeps a ledger of what has actually been verified, and refuses any claim the model makes that runs ahead of that ledger. The same system runs the appliances we ship. Results below are on Qwen 3.8 27B and Ornith 1.5 35B.

Policy-constrained tool use
15 of 20 tasks solved on a retail agent benchmark, graded on database state against ground truth — not on what the model claimed it did.
Real code, real tests
12 of 12 on a hard programming set, with every test re-run on a clean checkout.
Reliability
Zero malformed tool calls across the graded agent run.
Method
Three benchmarks in three separate domains, each layer switched on one at a time so its contribution is measurable, reported as paired per-task flips.
Models
Measured on Qwen 3.8 27B and Ornith 1.5 35B, both frozen. The weights are never touched — every gain comes from the scaffolding around them.

04 · Expert model & packs

Project Graft

A small expert model, plus parameter packs you swap in while it runs.

A graft joins living tissue to a host so it grows as one — a pack does that to a model, and can be removed as cleanly.

A compact model trained hard on reasoning, programming and mathematics — then extended on the fly by hot-swappable packs of parameters that carry a domain it was never trained on. Load a pack and the model gains the field; remove it and the knowledge goes with it. No fine-tuning run, no embedding model, no data leaving the building.

Clean-room by construction
The test material documents a software library invented for the experiment. No model on earth has seen it, so nothing can be explained by memorisation.
Controlled
A deliberately poisoned pack runs as a control arm — proving the model is reading the pack rather than guessing.
Why it matters
Domain expertise becomes a file you install and revoke, instead of a training run you cannot audit.

What investment unlocks

Planned,
not claimed.

Everything above has been measured. Everything below is roadmap, and depends on infrastructure we do not yet own. We keep the two apart on purpose.

Invest in Studio C
Next
A short recovery-training pass to lift low-bit retention further. The method is proven at small scale; running it properly needs compute we are yet to fund.
Then
Our current model line shipped pre-installed on a LocallyAI appliance — the compression and memory work is what makes that fit on hardware a business can buy outright.
Beyond
Larger models at the same footprint, and packs covering regulated domains.

Let’s build

Good systems.
Built together.

Tell us what you’re working on. We’ll help you move it forward.

Enquire now [email protected] More at LocallyAI.com.au

Investment

We’re opening the cap table. Register your interest and we’ll send you the detail.

Invest in Studio C