Bud Model Foundry
The sovereign-grade, multi-vendor, agentic-first training platform — build, fine-tune, post-train and agentic-train open models on your own infrastructure, with research-grade control and production-grade operations. No InfiniBand, no hosted dependency, no hardware tax.
Full-spectrum training, inside your perimeter.
Bud Model Foundry covers everything between raw data and a registered, production-ready model — data preparation, six training stages, step-level control, a built-in agentic-RL substrate, serving, and lineage — designed for sovereign deployment and engineered to perform on commodity hardware.
Layer 03 of the Bud stack.
Model Foundry is the training plane. Models trained here register with lineage and flow into Bud AI Foundry serving one layer up; production signal flows back down through drift detection and feedback.
Feeds upward — checkpoints register with full lineage and deploy into Bud AI Foundry serving at Layer 04, with no export/import seam. Training runs under Bud SENTRY governance like every other layer.
Builds on Bud Pod's pooled GPU compute and Bud LayerZero's kernel orchestration — which is how mixed NVIDIA, AMD, Qualcomm and Intel fleets appear as one training substrate.
The full training surface.
The full training surface, expanded to the specifics an evaluator needs — the training matrix, the DiLoCo mechanics, the Tinker primitives, the agentic-RL substrate, the data flywheel, and the five interfaces.
Full-spectrum training core
Most platforms cover a slice. Model Foundry covers the whole spectrum — all configurable, all validated.
Training stages 6
Fine-tuning methods 7
Quantization formats 9
Optimizer families 10
Every job launches with pre-flight memory, time and cost estimates per GPU — no OOM-then-restart — and streams loss, learning rate and GPU metrics live over WebSocket.
Bud DiLoCo — low-bandwidth distributed training
The capability that breaks the SXM dependency. Standard distributed training syncs gradients every step, demanding 100+ Gbit/s InfiniBand. Bud DiLoCo runs an inner AdamW loop per node, then syncs a pseudo-gradient through an outer Nesterov optimizer — cutting inter-node bandwidth by orders of magnitude.
Gradients exchanged at every optimization step — requires InfiniBand above 100 Gbit/s, locking you to SXM HGX hardware.
A single pseudo-gradient summarises hundreds of micro-updates — runs over standard Ethernet under 100 Mbit/s.
The 4,800× figure assumes int4 compression plus adapter-only synchronization and is a roadmap projection, not a shipping number.
Bud Tinker — step-level training control
Most platforms only let you submit-and-wait. Bud Tinker exposes the eight primitive operations of a training loop as REST endpoints and SDK methods — each call preserving full training state with bit-exact reproducibility, wrapped in the same auth, audit and encryption as a production pipeline.
Agentic RL & Simplified ART
Reinforcement learning is the dominant technique for agentic systems, and Model Foundry ships the entire substrate — not a generic RL toolkit. Three rollout modes (sync, async, streaming) run against real or simulated tools; ten graders, including a tool-call grader with partial-credit rewards, teach an agent to refine rather than just pass or fail.
Simplified ART layers a teaching metaphor over that substrate, opening agentic training to subject-matter experts, not just researchers. Five pre-built recipes — Reasoning, Code, Support, Tool Use and Safety — ship with sensible defaults.
Data pipeline & the continuous-improvement flywheel
Model quality is bounded by data quality, so the pipeline is a first-class subsystem: 260+ operators in a drag-and-drop DAG filter, dedupe, transform and balance data across text, image, audio, video and code. Pipelines run locally, multiprocess, or on a Ray cluster — the same DAG scales to TB-scale corpora — and every dataset version is rebuildable from its source and processing config.
Most AI initiatives stall because training and deployment are separate concerns. Model Foundry closes the loop inside the perimeter, with a full audit trail at every step:
Five interfaces to the same platform
From researchers writing Python, to operators in the dashboard, to autonomous agents calling Model Foundry as MCP tools — use whichever fits your team.
Sync + async, fluent builders, pre-flight estimates.
350+ endpoints, OpenAPI, webhooks, idempotency.
35-page GUI with progressive disclosure and 11 chart types.
Runs in any SSH session — air-gapped friendly.
Training as tools any LLM agent can call.
A job submitted through the SDK can be paused from the dashboard, monitored from the TUI, and registered by an MCP-driven agent — all on the same job, all with the same lineage.
Seven layers, each independently scalable.
Engineered for graceful degradation — core training and RL run on pure PyTorch and stay available even when optional high-level components are not.
| Layer | What lives there |
|---|---|
| Consumption | SDK, 350+ REST endpoints, 35-page dashboard, server TUI, OpenAI-compatible clients, MCP server. |
| Gateway | FastAPI with ordered middleware: request-ID, idempotency, size limits, rate limiting, auth, RBAC, CORS. |
| Execution | Celery workers (training, pipelines, imports) and in-process pipelines (Tinker, RL, fast inference). |
| Core engines | Bud Tinker, Training Pipelines, Bud RL Engine, Simplified ART, DiLoCo Orchestrator — all pure PyTorch. |
| Platform subsystems | Data pipeline, inference engine, model registry, drift detection, feedback collector. |
| Cross-cutting services | Auth, AES-256-GCM encryption, audit logging, cost tracking, notifications, idempotency. |
| Persistence | PostgreSQL for state + registry, Redis for cache/queues, MinIO/S3 for artifacts. |
One auth surface and one audit log span all seven layers — the same security envelope wraps a research-grade Tinker call and a production training pipeline.
One command, inside your perimeter.
No hosted dependency, no required outbound connection, no telemetry leaving your environment. Three deployment patterns from the same installer.
| Pattern | Best for | What you get |
|---|---|---|
| Single-node Docker Compose | Pilot & single-team production | Deployed in 30 minutes |
| Kubernetes via Helm | Multi-team production | Horizontal scaling |
| Air-gapped on-premise | Defence & classified workloads | Maximum sovereignty — no outbound path at all |
The security envelope
Every headline number, with its basis.
Each claim is paired with how it arises and the conditions it holds under — including the ones that are projections rather than measurements.
| Claim | Metric | Basis | Conditions & caveats |
|---|---|---|---|
| 100–500× bandwidth reduction | inter-node traffic | Sync cadence: pseudo-gradient per ~100 steps vs per-step gradients | 100× conservative · 500× typical multi-node |
| 4,800× bandwidth reduction | inter-node traffic | int4 compression + adapter-only sync | roadmap — not shipping |
| <100 Mbit/s sufficient | link bandwidth | Outer-loop sync over Gloo/Ethernet | vs 100+ Gbit/s InfiniBand conventional |
| 118+ supported models | architecture count | Platform support matrix | at release · grows upstream |
| 30 min deployment | wall-clock install | Single-node Docker Compose via bud-install | pilot pattern · production 30 min–days |
| Bit-exact reproducibility | training state | Tinker preserves full state at every primitive call | bit-identical pause / resume |
Counted surfaces — 6 stages, 7 methods, 9 quantization formats, 10 optimizers, 10 graders, 260+ data operators, 350+ REST endpoints — are inventory counts of shipping features, not benchmark results.
The same capability surface — on your hardware, inside your perimeter.
Against the three alternatives every evaluator has already considered: hosted fine-tuning services, hyperscaler ML platforms, and assembling the open-source stack yourself.
| Capability | Hosted services | Hyperscaler | DIY open-source | Bud Model Foundry |
|---|---|---|---|---|
| Sovereign / air-gapped | Fails | Partial | Pass | Pass |
| Predictable cost at scale | Per-token | GPU-hour + egress | CapEx + ops | License-based |
| Multi-vendor GPU support | Limited | Catalogue only | DIY | NVIDIA · AMD · Qualcomm · Intel |
| Agentic-RL stack built in | No | Partial | DIY | Full stack |
| Commodity-Ethernet training | No | No | Rare | Bud DiLoCo |
| Time to first production job | Days | Weeks | 6–12 months | 30 min – days |
| Lifecycle scope — registry · drift | Training only | Limited | DIY each | End-to-end |
Where the capability surface creates value.
The same platform, translated into concrete value across regulated and sovereignty-bound sectors.
Compliance-grade copilots
Compliance copilots with refusal training, audit-grade fraud-detection reasoning, and loan-origination assistants trained on internal policy.
Clinical reasoning, in-perimeter
Clinical reasoning on de-identified notes, vision-language radiology assistants, and federated training across hospital consortia.
Fully air-gapped agents
Intelligence-analysis agents with no outbound path, multilingual citizen-service agents, and cyber-defence reasoning models.
Sovereign AI PaaS, white-label
Sovereign AI Platform-as-a-Service, multi-tenant LoRA serving at scale, and cost leadership via commodity-Ethernet training.
Sweat the fleet you own
Production training on PCIe-cluster fleets, multi-node training over commodity Ethernet, and mixed NVIDIA + AMD scheduling.
National capability, national infra
National AI capability on national infrastructure — EU AI Act and DPDP provenance, reproducible datasets, and a complete audit trail.
The platform argument, in full.
This brief is the product reference. For the platform-level case — why training, serving, and governance belong on one plane — read the whitepaper, or see a training run land on your own hardware.
Platform White Paper
The Enterprise AI Management Platform
The authoritative narrative argument
Where the training plane fits the whole — the economics and governance case for running the full AI lifecycle on one sovereign stack.
Read the whitepaperBack to overview
Bud Model Foundry
Train open models on hardware you own
Return to the high-level product page — the five design commitments and the DiLoCo story at a glance.
Back to the product pagePut 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.