One sampled 27B server run: 4,361 tokens, MTP3 and correction on. A separate deterministic code task reached 92.1 tok/s. Separate single runs; decode only.
01 / THE OPPORTUNITYSTUDIOC · INVESTOR BRIEF
A 27B model.
In 9.03 GB.
StudioC makes capable AI practical on hardware a business controls. Smaller models, a native runtime and a private workspace, built together.
CAPABILITY RETAINED AFTER RECOVERY97.35%
Average retention across seven core benchmarks.
The original full-precision model is the reference.
Model files in decimal GB. Package size is separate from running memory.
Less hardware between a business
and useful AI.
Our starting point is private business AI: company knowledge, tools and approvals on a machine the customer controls. The ambition is infrastructure that makes capable AI practical far beyond the cloud.
Model-only thinking-on results. Peak memory for model + 32K context: 11.69 GB on Apple silicon, reproduced in 3 of 3 runs. Acceptance on a real 16 GB machine remains open.
About 3.3 A100 GPU-hours (US$5.30) for data capture, compression and recovery. Setup and data generation excluded. Full benchmark run: US$2.27.
Harry · Matt · Riley Founders / Hunter Region, Australia
Founded August 2026 in the Hunter Region, NSW, Australia.01 / 04
02 / THE MEASURED ADVANTAGESTUDIOC · INVESTOR BRIEF
Smaller weights.
Measured capability.
Longer thinking-on recovery brings the same 9.03 GB package closer to the original model. Native execution, memory and reliability are checked separately.
Selection set: down 63%. Independent confirm set: 0.625 to 0.254, down 59%. Lower is closer to the original model.
HumanEval answer accuracy, thinking enabled. Neither model uses the correction policy in this comparison.
| Benchmark | Retention |
|---|---|
| MBPP | 98.91% |
| HumanEval | 96.10% |
| GSM8K | 98.98% |
| MATH-500 | 93.98% |
| ARC-Challenge | 98.64% |
| MMLU | 96.18% |
| IFEval | 98.68% |
Thinking enabled, without correction. Mean retention: 97.35%; lowest core task: 93.98%. The reported twelve-task average is approximately 96.00%. Retention averages per-task score ratios, not answer accuracy.
| Task | Current | Original | Retention |
|---|---|---|---|
| MBPP | 91/100 | 92/100 | 98.91% |
| MATH-500 | 125/150 | 133/150 | 93.98% |
| HumanEval | 148/164 | 154/164 | 96.10% |
| GSM8K | 97/100 | 98/100 | 98.98% |
Correct answers / problems, compared with the full-precision original. These are current-model scores without correction. Overall corrected retention is not yet available for this model.
New-model output-limit cases: 13/164 HumanEval and 14/150 MATH-500, versus 27/164 and 26/150 previously. HumanEval repeated blocks: 3/164 versus the original model’s 1/164. The repetition check did not pass.
Fact retrieval / instruction constraints.
At 8K: 20/20 retrieval. Native tool tasks: 44/60.
Apple silicon, 3 of 3 runs. Acceptance on a real 16 GB machine remains open.
Measurement conditions & next acceptance gates +
At 32K, one tested attention operation fell from 38.96 ms to 2.165 ms, about 18× faster, not a whole-model speedup. The 4-bit cache uses about 3× less space than 16-bit at similar kernel speed.
The new divergence test uses thinking-on examples up to about 40K tokens, versus 1,024 previously. Raw values are not directly comparable between tests. Recovery still trains scales and norm weights, previously measured at 0.71% of parameters; that share was not recalculated. Codebook and indices are confirmed unchanged, at 2.69 bits per weight.
32K retrieval declined from 20/20 to 19/20, and constraints from 9/10 to 7/10. MMLU retention fell from 98.09% to 96.18%. Native tool-task score is 44/60 against the teacher’s 45/60: 97.78% retention. This differs from tool-format validity; the earlier 60/60 format check was not rerun.
An additional index optimisation failed its quality screen and is disabled (0.436 nats against a 0.02 limit). Windows and Linux production deployment are not yet accepted. Individual configurations are not independent repeat runs across model families.
Previous model: correction-policy results +
| Task | Alone | With policy | Retention |
|---|---|---|---|
| MBPP | 84/100 | 95/100 | 103.26% |
| MATH-500 | 115/150 | 134/150 | 100.75% |
| HumanEval | 135/164 | 151/164 | 98.05% |
| GSM8K | 94/100 | 96/100 | 97.96% |
Historical segmented A100 evaluation. The current model has not been rerun with this policy. Do not add these gains to its 97.35% core retention. The teacher reference was reused; only four of twelve planned tasks completed.
Previous-model output-limit counts fell from 27/164 to 1/164 on HumanEval and 26/150 to 5/150 on MATH-500. HumanEval repeated blocks rose from 0/164 to 2/164. These completion counts are optimistic relative to native enforcement. Above-100% task retention does not establish general superiority.
Founded August 2026 in the Hunter Region, NSW, Australia.02 / 04
03 / THE PRODUCTSTUDIOC · INVESTOR BRIEF
Private AI.
Real work.
For teams whose work lives in contracts, spreadsheets, email and company records. The Harness puts their knowledge, conversations and tools in one private workspace.
Compare supplier notices with your agreements.
The demonstration could not load. Please reload to try again.
Find it. Check it.
Cited answers and source passages. Compare agreements, analyse spreadsheets and turn meetings into owned actions.
Useful tools. Human control.
Reusable inline tools, scheduled reports and reviewed writes. Permission checks live outside the prompt.
Local, with doors.
StudioC Search needs no external search-provider API key. Mail connects directly to Outlook, Gmail and IMAP. Connect gives staff certificate-based access to the appliance.
Six Australian knowledge packs in development +
A national core and five specialist libraries alongside company records.
401source references
85workflow guides
60templates
- Australian national core
- Accounting & tax
- Construction & trades
- Healthcare & NDIS
- Legal & conveyancing
- Real estate & property
Federal · NSW · VIC · QLD · WA · SA · TAS · ACT · NT
Professionals make the final decisions. Packs support research and drafting. Qualified professionals verify current sources and make legal, clinical, tax, financial and compliance decisions.
Development inventory v2026.08.1. Jurisdiction selection includes the federal level and every state or territory; source coverage varies and current state-specific material is concentrated in NSW. Professional review and release signing remain pending. References may cover different sections of one authority.
Harness pilot preparation. Regular models are supported; installations with StudioC-compressed models are planned as hardware, security and deployment acceptance are completed. External connections stay optional.
Founded August 2026 in the Hunter Region, NSW, Australia.03 / 04
04 / THE COMPANY & WHAT COMES NEXTSTUDIOC · INVESTOR BRIEF
Start local.
Build much further.
Bring private AI into real businesses, then build on the infrastructure underneath it. A commercial path grounded in useful work, with room for a much larger company.
Machines, software, local experts.
LocallyAI AU plans appliance supply, installation and ongoing support across Australia through local experts. Start with a team’s actual workflow, a controlled pilot and clear acceptance checks.
Search is a business in its own right.
Exa and Tavily built businesses around search APIs for AI. StudioC Search brings search and page retrieval into our stack without an external search-provider API key. A standalone API, private hosting and professionally reviewed knowledge subscriptions are future opportunities alongside local installs.
All 65/65 software tests passed. Controls check malformed requests, repetition and unfinished responses.
Zero content violations in the replay. Release-runtime control outputs stayed identical: 4/4 with correction on.
An 8,192-token runaway was corrected to a right answer. The larger runaway test is running; no overall rate yet.
E-RLM / THE NEXT RESEARCH FRONTIER
Thinking that executes.
The model runs code during active reasoning, carries returned values into the next thought and lets the runtime assemble complete files at the end.
Two scripted demonstrations. Orange values and final files come from the runtime, never re-typed by the model. The execution protocol works; a general reasoning gain and reliable execution selection remain to be demonstrated on unseen problems.
Turn the technical result
into a dependable installation.
Next: 16 GB hardware at 32K, full-length reliability tests, secure Harness pilots and broader platform support. We’re building for the long term, from our home in the Hunter Region.
Team, progress & release scope +
Founded by Harry, Matt and Riley in August 2026. Current stage: recovered 27B release candidate, native Apple-silicon runtime and Harness pilot preparation.
All twelve correction components are integrated into the server. Software and recorded replay tests support the implementation; a broader live correction benchmark and independently repeated live results remain ahead.
Code availability is planned, with licensing still being decided. Future services require product validation and source-rights review; they would sit alongside local installations without moving local customers’ data.
Founded August 2026 in the Hunter Region, NSW, Australia.04 / 04