Skip to main content

​
OEM Bundling Framework — Pre-Configuration for Hardware SKUs

Date: 2026-04-01 Status: Planning Target Milestone: v0.24.0 Priority: P1 — drives OEM adoption of GAIA on Ryzen AI devices Related Issues:
  • #648 — OEM bundling framework: pre-configuration for hardware SKUs
  • #649 — Cost savings telemetry: show users what they save by running locally
  • #644 — Model Manager UI: download, switch, and update models from the desktop app
  • #615 — App Store and Distribution Channel Packaging
Related Plans:

​
1. Overview

AMD’s Ryzen AI 400-series is shipping across six major OEM partners: Acer, ASUS, Dell, HP, GIGABYTE, and Lenovo. Each partner ships devices with different memory configurations, NPU capabilities, and target audiences. The OEM bundling framework gives these partners a standardized way to pre-configure GAIA so that end users experience zero-touch deployment: plug in the device, and a private AI agent works immediately with no setup required. The first showcase device is the AMD Agent Computer (Halo Box), powered by Ryzen AI MAX+ 395 with 128GB unified memory.

​
2. OEM Configuration Manifest

Each OEM provides a YAML manifest placed at a well-known path during factory image creation. GAIA checks for this manifest before running its standard first-boot flow. Invalid manifests fall back to the standard gaia init flow with a warning. Required fields: schema_version, oem.name, hardware.sku, models.default. 
# oem-config.yaml -- OEM Bundling Manifest for GAIA
schema_version: "1.0"

oem:
  name: "ASUS"
  display_name: "ASUS AI Suite"
  support_url: "https://support.asus.com/ai"

hardware:
  sku: "ASUS-ProArt-PX13-HX370"
  processor: "AMD Ryzen AI 9 HX 370"
  npu_tops: 50
  gpu_compute_units: 12
  ram_gb: 32

models:
  default: "Qwen3.5-35B-A3B-GGUF"
  preinstalled:
    - id: "Qwen3-0.6B-GGUF"
      size_gb: 0.4
      purpose: "lightweight-chat"
    - id: "Qwen3.5-35B-A3B-GGUF"
      size_gb: 18.0
      purpose: "agent-reasoning"
    - id: "nomic-embed-text-v1.5-GGUF"
      size_gb: 0.3
      purpose: "rag-embedding"

skills:
  bundle: "consumer"    # consumer | developer | enterprise | creative
  enabled: ["chat", "rag", "talk", "summarize"]

settings:
  temperature: 0.7
  ctx_size: 32768
  first_run:
    skip_model_download: true
    skip_hardware_detection: true

​
3. Hardware Profiles

The framework maps hardware SKUs to model recommendations. OEMs select from predefined profiles or define custom ones in their manifest.
Profile TierRepresentative SKUNPU TOPSGPU CUsRAMMax Model SizeRecommended Default Model
Halo Max+Ryzen AI MAX+ 395 (Agent Computer)5040128 GB~200B (FP4)Qwen3.5-35B
HaloRyzen AI MAX 390503264 GB~70B (FP4)Qwen3.5-35B
HighRyzen AI 9 HX 370501232 GB~30B (Q4)Qwen3.5-35B
MidRyzen AI 7 34039816 GB~7B (Q4)Qwen3-8B-GGUF
EntryRyzen AI 5 32016416 GB~4B (Q4)Qwen3-4B-Instruct-GGUF
GAIA validates that a model fits the hardware at boot time. If the default model exceeds available memory (RAM minus ~4 GB OS overhead, with 15% model overhead for KV cache), GAIA falls back to the next smaller preinstalled model.

​
4. Pre-Installed Model Packages

OEMs bundle models onto the factory image or a companion USB drive. Three predefined tiers cover common deployment scenarios.
TierModels IncludedTotal DiskTarget Audience
MinimalQwen3-0.6B-GGUF~400 MBEntry-level devices, demo units
StandardQwen3-0.6B + Qwen3.5-35B + nomic-embed-text-v1.5~19 GBConsumer laptops (32+ GB RAM)
FullStandard + Qwen3-VL-4B + Qwen3-8B~25 GBDeveloper and Halo-class devices
Pre-installed models are placed at C:\AMD\GAIA\models\ and symlinked into the user’s ~/.gaia/models/ directory on first boot.

​
5. OEM Skill Bundles

Skill bundles define which GAIA capabilities are enabled out of the box. OEMs pick a predefined bundle (consumer, developer, enterprise, creative) or define a custom list in the manifest. Skills not in the active bundle are hidden from the UI but remain installed — users can enable additional skills from the Model Manager UI (#644) at any time.

​
6. Distribution Channels

ChannelPackage FormatAuto-UpdateTarget
Windows StoreMSIXVia StoreConsumer retail
Pre-installed imageEmbedded in OEM Windows imageVia GAIA update channelFactory install
USB provisioningZIP + PowerShell bootstrapManual / gaia updateIT deployment
Enterprise MDMMSI + Group Policy templateVia MDM policyFleet management
wingetwinget manifestVia winget upgradeDeveloper self-install

​
7. Cost Savings Telemetry

A key benefit of local AI is cost savings versus cloud APIs. The telemetry module (#649) tracks tokens generated and processed locally, then calculates equivalent cloud API costs using published pricing (OpenAI, Anthropic, Google). All data is stored locally in SQLite — nothing leaves the device unless the user explicitly opts into anonymous aggregate reporting. The Agent UI displays a savings summary card showing monthly and lifetime savings.

​
8. Phased Rollout

  • Phase 1 (v0.24.0): Ship the Agent Computer (Halo Box) with a production-ready reference OEM config, manifest schema v1.0, bootstrap loader in gaia init, cost savings telemetry v1, and Model Manager UI v1.
  • Phase 2 (v0.25.0): Provide OEM partners with a toolkit — gaia oem validate and gaia oem preview CLI commands, OEM integration guide, model packaging tools, and an automated test suite for manifest loading.
  • Phase 3 (v0.26.0+): Ship GAIA on all six OEM partner devices through standard distribution channels (Microsoft Store, MDM, winget), with enterprise fleet management tooling and a partner certification program.

​
9. GitHub Issue Cross-References

IssueTitleRelevance
#648OEM bundling frameworkPrimary issue — this plan is the design document
#649Cost savings telemetrySavings calculator and dashboard
#644Model Manager UIHardware-aware model recommendations
#615App Store and Distribution Channel PackagingMSIX, winget, Store distribution
#597Setup wizardFirst-boot flow that OEM manifest can skip or customize
#612Agent registrySkill bundles map to registered agents
#643System trayOEM branding in tray icon and tooltip
#634Always-on agent (heartbeat)Enterprise fleet health monitoring
#474MCP auto-installEnterprise MCP server provisioning via manifest