Ancapex Research builds the training, evaluation, provenance, and ownership layer for community-owned AI models. The missing layer between decentralized GPUs and sovereign models.
Intelligence is not rented.
Compute is not enough.
Models must be trained, verified, and owned.
Inference is the bridge.
Training is the territory.
Ownership is designed before the first serious run.
Available GPUs are not enough. Training demands data, fault tolerance, reproducibility, evaluation, and clear ownership of every weight and checkpoint.
A different machine: memory, bandwidth, checkpoints, determinism, failures, and reproducibility — not just serving a model.
Provenance, licensing, privacy, contamination checks, and eval splits decide what a model can become.
Weights, checkpoints, recipes, datasets, and rights must have explicit owners before financialization begins.
A decentralized training network must verify work from untrusted or unstable compute contributors.
Inference is useful. But the deeper prize is the ability to form, improve, audit, and own models outside centralized labs.
“Gonka runs model inference.”
“A decentralized network can train, verify, and own models — not just rent access.”
A staged path from open-weight sovereignty to a model trained materially on decentralized compute. Each phase ships real artifacts before the next begins.
Start with Kimi K2.6-class open weights: self-hosted deployment, refusal transparency, behavior maps, and sovereign alignment.
Reproduce and extend Prime Intellect-style post-training: SFT, RL, verifiers, environments, and agentic rollouts.
Turn decentralized GPUs into a training network: qualification, scheduling, checkpointing, and work verification.
Lineage for every model: configs, dataset hashes, checkpoints, evals, regressions, and reproducibility.
A refinery for model-grade data: licensed datasets, synthetic tasks, verifiable environments, and reward functions.
Design ownership before tokenization: contributors, data owners, researchers, and rights-holders get explicit claims.
The long-term target: a frontier-adjacent model trained materially on decentralized compute — transparent end to end.
No fake training. Every serious claim must leave a trail.
Configs, loss curves, hardware, run duration, failures, and reproducibility notes.
Capabilities, limitations, refusal behavior, eval results, and deployment assumptions.
Where a model refuses, where it complies, and whether the boundary is explicit or hidden.
GPU type, VRAM, uptime, bandwidth, interconnect, node churn, and failure rates.
Domain tests, regression tracking, contamination checks, and honest limits.
Failed experiments stay visible. Failure is part of the research record.
Sovereignty means inspectable behavior, explicit boundaries, and owned weights.
Every run leaves config, checkpoints, logs, metrics, and evals.
Provenance, licensing, privacy, and contamination checks are explicit before training.
Real domain tests and honest limits, not leaderboard theater.
Refusal behavior should be explicit, testable, and documented.
Weights, checkpoints, recipes, datasets, and rights have clear owners.
Pipelines, evals, checkpoints, and provenance come before any story about the network.
Bring compute, data, models, or capital. We bring the training stack, evaluation discipline, and ownership design.
Thank you. We respond directly to serious counterparties with a credible fit.
REF · —