Anneal
Compress existing models. Recover capability through training. Compare against the original.
01 / THE COMPANYSTUDIOC · INVESTOR BRIEF
Private AI.
Worth owning.
StudioC makes capable AI practical on hardware a business controls — through smaller models, faster local infrastructure and a private workspace.
RECOVERED MODEL27B
8.48 GB model file · Established build
BENCHMARK RETENTION GATES5/5
Above 80% of teacher scores · Matched tests
CORRECTED HARNESS TOOL CALLS300/300
287/300 before correction · Unchanged weights
Established model · GSM8K 91.9% · MMLU 97.1% · HellaSwag 100.6% · ARC Challenge 92.0% · HumanEval 81.4%
The model. The engine. The workspace.
One private AI stack.
rMLX
Our Rust and Metal runtime. The graph, kernels and serving path built for local hardware.
The Harness
A private workspace for company knowledge, cited answers and tools that do useful work.
E-RLM
Our next research frontier: execution inside active reasoning. Protocol working; capability gains still to prove.
Search can be a business of its own.Companies such as Exa and Tavily have built businesses around search APIs for AI. StudioC Search puts search and page retrieval inside our stack, without requiring a key to an external search API. A standalone StudioC Search API is a future commercial opportunity.
One stack. Several routes to revenue.Local installations and support create the first customer relationship. Dedicated private hosting is planned. Search APIs and specialist knowledge subscriptions could create additional recurring revenue. These future offers would sit alongside local installs: no change to their operation and no move of customer data.
Founder-funded by Harry, Matt and Riley. Pre-seed funding advances business pilots, runtime integration and research. Harness currently uses stock models. Future standalone services would need product validation and source-rights review.
Founded August 2026 in the Hunter Region, NSW, Australia.01 / 04
02 / ANNEAL + rMLXSTUDIOC · INVESTOR BRIEF
Smaller weights.
Serious capability.
A 27B model compressed to 8.48 GB, retaining more than 80% of teacher performance across five matched benchmarks. The established recovered build passes all five retention gates.
| Benchmark | Original correct | Established · 12 Sep | Newer · 13 Sep |
|---|---|---|---|
| GSM8K · maths | 246/250 | 226/250 · 91.9% | 222/250 · 90.2% |
| MMLU · knowledge | 479/570 | 465/570 · 97.1% | 461/570 · 96.2% |
| HellaSwag · commonsense | 685/1000 | 689/1000 · 100.6% | 693/1000 · 101.2% |
| ARC Challenge · science | 363/500 | 334/500 · 92.0% | 334/500 · 92.0% |
| HumanEval · code | 129/164 | 105/164 · 81.4% | 94/164 · 72.9% |
Each cell shows correct answers and retained teacher score. Established build: 5/5 gates above 80%; newer variant: 4/5, with coding recovery in progress. GSM8K: 8-shot, strict match, 1,024-token cap. MMLU: 5-shot. HellaSwag: 10-shot; ARC: 25-shot, normalized accuracy. HumanEval: unfiltered pass@1, 1,024-token cap. Matched subsets, materialized BF16 on CUDA; separate from native execution.
NATIVE EXECUTION / M1 MAX · METAL
16.58tok/s
At 4K context · 64 generated tokens
Newer recovered variant · One measured run
125.26 tok/s prompt processing · 24.07 s load.
8.48 GB model file · 10.707 GiB resident weights · 11.936 GiB peak RSS. Memory and long-context optimisation remain in development; sub-10 GB total runtime is not yet achieved.
Metal numerical check: mean forward KL 0.12324, argmax agreement 84.18%, 1,024 positions across four contexts. Passed its comparison gate; separate from task accuracy.
| Paired workload | Speed ratio |
|---|---|
| Graph building | 3.07× |
| Dispatch overhead | 1.90× |
| Fused 2,000-operation chain | 7.95× |
| 64 MB tanh-GELU activation | 6.48× |
| fp16 4096³ matrix multiplication | 1.00× |
| Decode microbenchmark | 0.93× |
Paired medians, n=14; 174/174 parity cases. Component gains are separate from full-model speed. The decode microbenchmark is 408.5 vs 437 tok/s.
Founded August 2026 in the Hunter Region, NSW, Australia.02 / 04
03 / THE HARNESSSTUDIOC · INVESTOR BRIEF
From a question.
To work done.
Private AI that can actually do work. Company knowledge, tools and approvals in a workspace deployed on hardware the business controls.
Find. Make. Follow through.
Cited answers and source passages. Reusable inline tools and spreadsheet charts. Meeting transcripts, owned actions and scheduled reports on your machine.
StudioC Search, Mail & Connect.
Web search without an API key to an external search provider. Direct mail integration with Outlook, Gmail and IMAP hosts. Certificate-based remote access to your appliance.
Industry context. Company control.
Versioned source references, workflows and templates alongside company files. Read-only, approved writes or broader access per chat. External connections stay optional.
Knowledge worth building on.
Six packs: one national core and five specialist libraries. Federal context, with every state and territory available in the jurisdiction settings.
401source references
85workflow guides
60templates
Federal · NSW · VIC · QLD · WA · SA · TAS · ACT · NT
Australian national core
Privacy, consumer law, employment records and shared business obligations.
Construction & trades
Contracts, licensing, WHS, variations and payment claims, with NSW guidance.
Healthcare & NDIS
Privacy, consent, clinical records and provider obligations.
NSW legal & conveyancing
Conveyancing law, client identity, costs disclosure and settlement checks.
Accounting & tax
ATO guidance, GST/BAS, payroll, super and practice records.
NSW real estate
Agency and tenancy rules, disclosure, trust money and property management.
Built to support professionals. Professionals in their respective fields verify current sources and make final legal, clinical, tax, financial and compliance decisions. The packs support research and drafting; they do not replace professional judgement.
Development inventory, v2026.08.1. Source coverage varies by field and jurisdiction; current state-specific material is concentrated in NSW. Professional review and release signing remain pending. References can cite different sections of one authority.
287 → 300/300
Automatic correction of malformed structured tool calls — without changing model weights. Same 300 prompts, same stock 4-bit model.
12/12
Exercism-based code set. Tests rerun on a clean checkout. A separate, small internal test.
Start with the workflow.Closed pilots in preparation. LocallyAI AU handles the appliance, on-site setup, accounts, groups and support. Trained contractors are the planned route to wider delivery.
A clear release path.Code availability and free personal/business use are planned; licence terms remain in development. Embedding in another product will require approval and may carry a fee.
Founded August 2026 in the Hunter Region, NSW, Australia.03 / 04
04 / E-RLM + WHAT COMES NEXTSTUDIOC · INVESTOR BRIEF
Thinking that
executes.
Our next research frontier. E-RLM explores a different model/runtime boundary: deterministic execution inside active reasoning, with returned values carried into subsequent thinking and full files assembled at the end.
The original two scripted demonstrations. Orange values and final files come from the runtime, never re-typed by the model.
RESEARCH FRONTIER / CAPABILITY PROOF PENDING
A working protocol. The next proof ahead.
The execution protocol and runtime are working. The current fine-tune has learned the execution notation, but has not yet demonstrated improved problem-solving or reliable execution selection.
Next proof
Held-out problems → execution selection → correct answers → clean stopping → generalisation.
A foundation we can test.The Rust controller operates inside decoding; an isolated Python/SymPy worker executes operations. Both hosts passed 61 runtime gates and eight focused recovery tests. The original demonstrations show the intended interaction, not a reasoning benchmark.
Fund the platform. Advance the frontier.Business pilots and compressed-runtime integration lead the next product milestones. E-RLM follows an explicit capability gate: demonstrate useful, reliable execution on unseen problems before making a product-performance claim.
13 September 2026. Training completed: 141 steps, six epochs, 110 corpus replays; validation NLL 2.257 → 0.682. These are process measurements. Initial matched pilot: maths 7/8 versus base 8/8; reasoning 2/4 versus 4/4; neither selected execution. Small sets, full campaign incomplete. Earlier validation contamination and conflicting training targets prompted data repairs; the measured adapter has not established a reasoning gain.
Founded August 2026 in the Hunter Region, NSW, Australia.04 / 04