Train open models on the hardware you already own.
Build, fine-tune, post-train and agentic-train open models on your own infrastructure — research-grade control, production-grade operations, and no hardware tax. Bud DiLoCo trains across nodes over commodity Ethernet, so the SXM-and-InfiniBand dependency ends here.
The whole training spectrum, one platform.
Most platforms cover a slice of the training lifecycle. Bud Model Foundry covers all of it — data preparation, six training stages, step-level control, a built-in agentic-RL substrate, and a registry with lineage — behind one auth surface and one audit log, deployed inside your perimeter.
- Bud Model Foundry is Layer 03 of the eight-layer Bud stack — the training plane: the whole training spectrum on one sovereign platform, 118+ open models across 4 GPU vendors, deployed inside your perimeter.
- Bud DiLoCo: conventional distributed training syncs gradients every step and demands 100+ Gbit/s InfiniBand; DiLoCo runs an inner AdamW loop per island and syncs one pseudo-gradient through an outer Nesterov optimizer every ~100 steps — 100–500× less inter-node bandwidth, standard Ethernet under 100 Mbit/s.
- Six capability groups: full-spectrum training core; Bud DiLoCo; Bud Tinker; agentic RL & Simplified ART; data pipeline & flywheel; five interfaces.
- The result: your models, your hardware, your perimeter — no hardware tax.
Full-spectrum training, on hardware you already own.
Everything between raw data and a registered, production-ready model — data preparation, six training stages, step-level control, a built-in agentic-RL substrate, and lineage — designed for sovereign deployment and engineered to perform on commodity hardware. No InfiniBand, no hosted dependency, no hardware tax.
Five commitments, built into the foundation.
Each one is a structural answer to a specific failure of the existing market — designed in, not bolted on. The sixth is the capability that breaks the hardware tax.
Commodity-Ethernet training
Bud DiLoCo islands train locally and sync one pseudo-gradient — no per-step gradient exchange, so the SXM-and-InfiniBand dependency ends here.
Sovereign by deployment
One-command install inside your perimeter — no outbound dependency, air-gapped as a first-class pattern, AES-256-GCM at rest.
Multi-vendor by design
NVIDIA, AMD, Qualcomm and Intel, with PCIe cards as first-class hardware — mixed fleets schedule into a single training job.
Agentic-first by purpose
A full RL substrate ships in the box — and Simplified ART's teaching metaphor opens agentic training to domain experts, not just ML researchers.
End-to-end by scope
Data prep, training, RL, registry with lineage, and drift detection on one platform — every dataset version rebuildable, the loop closed inside the perimeter.
Production-grade from day one
bcrypt-hashed API keys, OAuth/OIDC, model-level RBAC, atomic quotas, four rate-limiting algorithms, structured audit, Prometheus metrics.
What's inside the foundry.
The training plane is built from named engines — the distributed-training orchestrator, the step-level control surface, and the substrate that trains agents.
Bud DiLoCo
An inner AdamW loop per island, one pseudo-gradient through an outer Nesterov optimizer every ~100 steps — same total step count, a fraction of the traffic.
Bud Tinker
The eight primitive operations of a training loop as REST + SDK — full training state preserved at every call, for custom RL and step-level debugging.
Simplified ART
Student, coach, curriculum, grader — a teaching metaphor that compiles to a full RL run, with pre-built recipes for reasoning, code, support, tool use, and safety.
The full story, in depth.
The training matrix in full, the DiLoCo mechanics, the Tinker primitives, the ART pipeline — and every headline number paired with how it was measured.
Product Brief
Bud Model Foundry Product Brief
The deep-dive product reference
The full training matrix, Bud DiLoCo step by step, the eight Tinker primitives, the Simplified ART pipeline, the seven-layer architecture, and the comparison table.
Read the product briefPlatform White Paper
The Enterprise AI Management Platform
Where the training plane fits the whole
The platform-level argument — why training, serving, and governance belong on one plane, and the economics that follow when they do.
Read the white paperPut your data on it.
The fastest way to see what an integrated AI operating system does for your enterprise is a proof-of-concept on your infrastructure, with your data.