Studio KRiX
Portal
Menu

Studio KRiX · Connected Systems

Local intelligence,
closer to the systems it supports.

Local AI brings language models, reasoning and automation closer to the systems and data they support. Selected workloads can run locally on dedicated hardware instead of sending every prompt, document, household event or diagnostic log to a cloud service.

Explore the architectureConnected Systems

Representative hardware

Two high-memory paths

These platforms represent two ways to bring substantial model capacity into a compact local system. They are options under consideration, not claims about hardware currently owned or deployed by Studio KRiX.

NVIDIA DGX Spark

GB10 Grace Blackwell Superchip

  • 128 GB coherent unified memory
  • 273 GB/s memory bandwidth
  • Up to 1 PFLOP FP4
  • CUDA and the NVIDIA AI ecosystem

A compact Arm-based system suited to large local language models, coding agents, multimodal work and retrieval-augmented generation. Its unified memory can accommodate very large quantised models when the model, context and runtime fit within the available pool.

Representative model fit
  • gpt-oss 120B
  • Larger Qwen3 variants
  • Multiple smaller models
  • Multimodal workloads
  • Coding agents
  • RAG pipelines
NVIDIA specifications (opens in a new tab)

AMD Ryzen AI Max+ 395

Ryzen AI Halo-class platform

  • 16-core Zen 5 CPU
  • Radeon 8060S integrated graphics · 40 CUs
  • Up to 128 GB unified LPDDR5x
  • Approximately 256 GB/s memory bandwidth
  • Linux acceleration paths through ROCm and compatible runtimes

A compact x86 platform with a large shared memory pool, attractive for familiar desktop and development workloads as well as high-memory local inference.

Representative model fit
  • Qwen3 8B–32B class
  • Gemma 3 12B / 27B
  • gpt-oss 20B
  • Devstral 24B
  • Larger quantised-model experimentation
AMD Ryzen AI Halo specifications (opens in a new tab)
Model fit is workload-specific.

Practical fit depends on quantisation, context length, runtime overhead, allocated memory, concurrency and model architecture. These are representative pairings, not speed or capacity guarantees.

Local model runtime

Models first, with a local serving layer

Qwen3, Gemma 3, gpt-oss and Devstral represent different kinds of work across reasoning, multimodal interpretation and software engineering. A restrained runtime layer can manage and serve compatible models to Lakaz, Connected Systems and development tools without becoming the public-facing capability.

Local runtimes such as Ollama can provide model management and serving.

  1. AI HARDWARE

    High-memory NVIDIA or AMD platform

  2. LOCAL MODEL RUNTIME

    Model management · Model serving

  3. LOCAL MODELS

    Qwen3 · Gemma 3 · gpt-oss · Devstral

  4. LAKAZ / CONNECTED SYSTEMS / DEVELOPMENT TOOLS

    Approved household · systems · engineering workflows

A conceptual runtime path. The exact services, models and integrations depend on the hardware, operating system and approved use case.

Representative local models

Concrete models, chosen for the work

Representative models suited to this architecture include the families below. They are candidates, not a claim that Studio KRiX currently runs every model or size listed.

Representative local AI models, their roles and example uses
ModelRoleExample use
Qwen38B · 14B · 30B · 32BGeneral reasoning / agentsAssistant work · reasoning · multilingual interaction · tools · agentsLakaz assistant · tool use · multilingual interaction · agent workflows
Gemma 34B · 12B · 27BVision + textMultimodal reasoning · document and image interpretationApproved image understanding · document interpretation · multimodal workflows
gpt-oss20B · 120BHeavy reasoningStructured reasoning · agentic workflows · larger local inferenceComplex analysis, agentic workflows
Devstral24BSoftware engineeringCoding · codebase exploration · multi-file work · development agentsLocal coding agents, diagnostics tooling

Optional intelligence service

Local AI + Lakaz

Local AI could sit behind Lakaz as a focused intelligence service: interpreting approved context, finding relevant information and proposing useful summaries or next steps.

Explore Lakaz
  1. Natural-language household queries
  2. Household and system status summaries
  3. Task extraction
  4. Maintenance assistance
  5. Documentation and manual RAG
  6. Network-log summaries
  7. Automation suggestions
  8. Event prioritisation
  9. Local voice intent interpretation
  10. Camera-event description where explicitly enabled
  11. Household knowledge search

Guarded architecture

Intelligence does not replace permission

The AI does not receive unrestricted authority over critical systems. It can interpret, retrieve, summarise and suggest; Lakaz remains responsible for deterministic permissions, confirmation and audit.

Critical actions require explicit approval
  • Unlocking doors
  • Changing access permissions
  • Disabling alarms
  • Modifying firewall rules
  • Exposing cameras
  • Changing critical infrastructure

  1. Data & events

    • Documents
    • Tasks
    • Sensors
    • Network logs
    • Approved imagery

  2. Runtime & model management

    • Local model serving
    • Model management

  3. Models

    • Language
    • Vision
    • Embeddings
    • Coding

  4. Lakaz intelligence layer

    • Context
    • Retrieval
    • Summaries
    • Suggestions

  5. Policy & permissions

    • Allowed actions
    • Confirmation
    • Audit

  6. Household / user

    • Clear information
    • Explicit choices
    • Human approval
Conceptual architecture only. No private live data or unrestricted system access is represented.

Privacy & security

Local by design, secure by practice

Local inference can create clearer data boundaries, but locality alone is not a security control. The complete system still needs careful maintenance, isolation, authentication and permissions.

  1. Sensitive information can remain on local infrastructure
  2. No cloud API required for selected workflows
  3. Local document retrieval
  4. Predictable data boundaries
  5. Potentially lower recurring inference cost
  6. Usable during some internet outages

Security dependencies

  • OS patching
  • Network isolation
  • Authentication
  • Service configuration
  • Model and tool permissions
  • Backups
  • Audit logs

Local AI is not inherently secure. Its safety depends on how the hardware, operating system, runtime, models, tools and network are configured.

Part of Connected Systems