Run a model
Inference uses existing model weights. Memory, quantization and context length determine what fits.
DIY AI / CONCEPT SERIES
Your models.
Your hardware.
Install compatible open models, build an agent, connect a camera or explore fine-tuning. Start with a modular platform that matches what you want to learn and build.
Concept configurations for co-development. These are not stocked products, validated AiGain kits or published reference designs. Final components, model compatibility, price and delivery are agreed after a feasibility review.
Adapt a model
Fine-tuning adjusts a pretrained model using your data. LoRA/QLoRA can reduce the resources needed.
Train from scratch
Small learning experiments are possible. Large foundation-model pretraining is a separate, much larger infrastructure project.
LOCAL INFERENCE + AGENT DEVELOPMENT
Agent Box
A modular desktop starting point for local assistants, document search and tool-using agents.
- Hardware direction
- Proposed starting configuration: Linux-compatible x86 mini PC, 32–64 GB RAM, 1 TB NVMe SSD and optional GPU expansion on a suitable chassis.
- Software direction
- Evaluate Ollama or llama.cpp with a compatible, appropriately licensed model. Add an agent runtime and explicitly approved tools.
- What to validate
- CPU-only performance depends on model size and context. This is primarily an inference and development concept, not a large-model training system.
- Customization
- Memory, storage, enclosure, deployment image and documentation.
CAMERAS + LOCAL AI INFERENCE
Edge Vision Kit
An accessible prototype platform for vision experiments, connected sensors and robotics peripherals.
- Hardware direction
- Candidate platform: NVIDIA Jetson Orin Nano Super developer kit with 8 GB shared memory, NVMe storage, a compatible camera and active cooling.
- Software direction
- Evaluate supported JetPack software, a suitable vision model and the device interfaces required by the application.
- What to validate
- Inference is the main target. Larger-model training should run on separate GPU hardware; trained models can then be evaluated on the edge device. No benchmark has been measured for this AiGain concept.
- Customization
- Camera, lens, enclosure, mounting, sensors and connectivity.
SMALL DEVICES + HANDS-ON LEARNING
Maker AI Node
A DIY node for learning local AI, sensing and device integration with replaceable modules.
- Hardware direction
- Candidate platform: Raspberry Pi 5 with a compatible AI HAT+ 2, suitable power supply and cooling. The HAT provides its own 8 GB memory.
- Software direction
- Use the supported Hailo software and compatible models. This is a separate software path from a general-purpose GPU workstation.
- What to validate
- The accelerator targets inference. It does not run every open model, and is not proposed as a general-purpose training accelerator.
- Customization
- Sensors, camera or audio peripherals, housing and learning documentation.
LOCAL INFERENCE + PARAMETER-EFFICIENT FINE-TUNING
Personal AI Lab
A serviceable workstation concept for developers who want to experiment with their own datasets and model adaptations.
- Hardware direction
- Proposed starting configuration: a supported GPU with 24–32 GB VRAM, 64–128 GB system RAM, 2 TB NVMe storage and a power/cooling design sized for the final components.
- Software direction
- Evaluate PyTorch, Transformers and PEFT for LoRA/QLoRA workflows, with version-compatible drivers and a tested environment.
- What to validate
- Feasibility depends on the exact model, sequence length, batch size and training method. This is not a promise to pretrain a large foundation model. Small educational models can be scoped separately.
- Customization
- GPU and memory options, quiet cooling, assembly, environment setup and reproducible training examples.
A useful DIY package
An agreed package can include a bill of materials, assembly guide, wiring and mechanical files, installation steps, a small test dataset and a validation checklist. Proprietary module internals and third-party licenses remain outside any blanket open-source promise.
Platform references
- NVIDIA Jetson Orin Nano Super
- Raspberry Pi AI HAT documentation
- Hugging Face PEFT and quantization
- Ollama tool support
References checked September 2026. Platform names identify candidate components and do not imply a partnership or endorsement.
Build the next step with us.
Share your target model, workload, budget range and the part you want to build yourself. We can scope hardware, assembly and technical support around it.