Not raising. Open to the right conversation.
Research & prototype stage · Inbound only
MegaCpp is at the research and prototype stage, fully funded so far by DataSunrise Ltd (USA). We are not in fundraising mode and are not running a formal process. We are, however, open to conversations with aligned long-term investors who see the specialist-ensemble thesis and reach out directly — on the occasions when outside capital is strictly additive to what we're already building.
Research and prototypes, with a deliberate runway to production
Funding & structure
Public operations run through Datasunrise OÜ (Tallinn, Estonia), the brand operator. All research, training runs, evaluation infrastructure, and public writing to date have been funded by DataSunrise Ltd (USA), the parent company. There are no outside investors yet.
Stage
Research and prototypes. The architecture is designed and documented, multiple proof-of-concept systems are running end-to-end, and the technical blog records the experiments as they land. The parent company finances the runway. The only reason we'd take outside capital is to move faster than that runway allows — and only with a partner who makes the work better, not just the balance sheet.
What we've proved so far
The stack already incorporates the techniques the frontier field has converged on — and a few non-obvious ones we chose specifically for the C++ target. Every item below is backed by a public article or a reproducible sample; we prefer receipts to slides.
Hybrid architecture (NAM56R)
Mamba 3 plus Transformer experts, 16 routed experts with top-4 activation. 4.73B total, 3.03B active parameters — a compact design chosen for C++ workloads, not scaled for general chat.
Production-grade systems choices
NVFP4 inference and FP16 training, Muon optimizer, block-sparse attention, context and sequence parallel, activation-checkpointing policies. Documented in technical writing, not claimed in slides.
POCs with receipts
Ablation gates at 10K steps, FIRE/DASH/ReDo plasticity toolkit, STP geodesic regularizer, evaluation harnesses, TPU and Blackwell paths. The evidence is public: each claim links to the article and the reproducible sample.
Compute-aware design
The ensemble is built to run on the silicon the next generation of developer tools will actually afford — consumer and datacenter Blackwell, TPU v5/v6 — instead of assuming unlimited frontier-scale training budgets.
Why specialist ensembles beat general-purpose frontier models on engineering work
1 · Frontier labs are compute-bound
Every frontier provider is visibly capacity-constrained. Each incremental capability they ship competes for the same training and inference silicon. That ceiling is a structural opening for teams who design for a specific domain and keep parameter budgets, precision, and routing disciplined.
2 · General-purpose models dilute focus
Generalist models optimize for the average of every domain at once. That erodes the signal on concrete engineering skills — cross-file edits, build-system-aware reasoning, ABI-safe refactors, template-heavy C++. A specialist ensemble that trains on the real shape of engineering work can out-reason a 10-100× larger generalist on the tasks that matter most in developer workflows.
MegaCpp is a concrete bet on that gap. C++ is a large, high-value engineering domain with its own compiler, build, ABI, and review surface. A compact, compute-aware specialist ensemble that actually understands that surface is a real breakthrough — and a defensible product category, not a feature other people can ship tomorrow.
If the thesis lands after reading, say hello
We're not running an outbound process and don't work with placement agents. The conversations that move fastest are with investors who have read the architecture and a few articles first, then reach out with a short note on their thesis fit and why specialist AI infrastructure matches their current deployment window. Email to either founder lands fastest.
We reply to substantive notes within a few business days. Generic intros, deck requests, and placement-agent outreach are filed without reply. We're happy to talk even when nothing is imminent — keeping the right people in context is part of how a small research team makes good decisions when the moment is right.
Important notice
This page is informational only. Nothing on it constitutes an offer to sell or a solicitation to buy any securities, an investment recommendation, or a commitment of any kind, and the company is not presently conducting a securities offering. Any offer, if and when made, would be extended solely through definitive offering documents delivered to qualified investors in permitted jurisdictions. MegaCpp is a research initiative operated by Datasunrise OÜ (Estonia) and funded by DataSunrise Ltd (USA). Statements about the technology, its capabilities, and future direction describe current and intended engineering work; they are not projections of financial performance.