Keep what matters.
Anneal makes existing models smaller, then trains a small part of the model to recover useful capability.
01 / CAPABLE AI. CLOSER TO YOU.THE STUDIOC STACK
Big capability.
Within reach.
Running useful AI should not mean handing over your company’s knowledge or needing a room full of hardware. We work on the whole system: making models smaller, bringing their capability back, and giving people a private place to use them.
Built for the machine.
Our native runtime executes compressed models on Apple silicon. Reliability controls check how the model responds and uses tools.
The Harness.
A private workspace for conversations, company files, research and tools. Useful work, with permissions set by people.
02 / SMALLER. STILL CAPABLE.15 SEPTEMBER 2026
More of the model.
Less of the weight.
Our latest 27-billion-parameter release candidate fits in a 9.03 GB package. After recovery, it averages 97.35% retention of the original model’s scores across seven core benchmarks.
9.03GB27B model package
97.35%Core capability retained
32KContext evaluated
Model files, in decimal GB. Package size is not total running memory.
Original teacher = 100%. Average of each task’s retention, without realtime correction. Retention is separate from answer accuracy.
Measured peak: 11.69 GB for the model plus 32K context on Apple silicon, reproduced in 3 of 3 runs. Package size is separate from running memory. Acceptance on a real 16 GB machine remains open.
Explore the results03 / MORE CAPABILITY. DELIVERED.MODEL & RUNTIME, MEASURED
Good thinking.
A useful finish.
A model can know what to do and still get stuck repeating itself or run out of room before answering. Our runtime controls target those failures, helping more answers reach completion.
164 problems, thinking enabled. Both models run without the correction policy. These gains come from the updated model.
CURRENT MODEL. BEFORE CORRECTION.
More answers scored correct.
HumanEval improved from 135/164 to 148/164 correct with the new recovery. The current model also scores 91/100 on MBPP, 125/150 on MATH-500 and 97/100 on GSM8K, all without the correction policy.
Twelve correction components now run inside the server. All 65/65 software tests passed; replaying 1,876 recorded generations found zero content violations. These checks are separate from model benchmark accuracy.
Explore the controls and benchmarksWith correction enabled, 4/4 control outputs stayed identical. One live runaway was corrected to a right answer; a larger test continues. Overall retention with correction is not yet available for the current model.
04 / FROM THE LAB TO YOUR WORKDAY
Your work.
Your machine.
Find answers in your records. Work with your inbox. Build a chart, research a supplier or turn a meeting into next steps. The Harness brings it together, with sources and actions in view.
Know where it came from.
Company files and cited source passages, alongside six Australian knowledge packs in development.
Open the doors you need.
StudioC Search needs no external search API key. Mail connects to your email host. Connect brings remote staff back to your appliance.
People set the permissions.
Choose read-only work or approved actions. Keep local work on your hardware and external connections optional.
Pilots and deployment acceptance are in progress. Installs are planned with regular and StudioC-compressed models, selected for the machine and the work.