AI directly on selected Android phones.
8–24 GB RAM
Small local models can run on modern phones. More RAM gives more headroom. Current examples include the OnePlus 13 with 12 or 16 GB RAM and ASUS ROG Phone 9 Pro variants with 16 or 24 GB RAM.
Local AI deployment
Run useful AI on equipment you control.
Local AI means selected models run on your own phone, computer or private server instead of sending every request to a remote cloud service. Talunza can assess the workload, match suitable hardware and software, install the system and make it usable for the people who need it.
Where it fits
Local does not mean disconnected.
A school, lodge, business or home can keep Starlink or fibre and still run AI locally for privacy, resilience, controlled access and fast service over the local network. Some deployments are fully offline; others use cloud AI only when it adds value.
Private AI on a laptop or desktop.
16–64 GB RAM
Suitable for personal assistants, local document search, coding help and light office use. A discrete GPU is optional for small models but can improve speed significantly.
Larger models on dedicated graphics hardware.
8–32 GB+ VRAM
Examples range from 8 GB laptop GPUs for selected mid-size quantized models to 24 GB cards such as the RTX 4090 for larger local deployments.
One local AI server for many approved users.
Sized to workload
Schools, lodges, farms and organisations can share a controlled AI service over LAN or Wi-Fi. High-memory systems can support larger models and more concurrent users.
For schools and parents
Give learners AI without making the open internet the default for every question.
Starlink can transform a school, but abundant connectivity also increases the need for digital-safety controls. A local school AI server can provide an approved learning assistant over the school network while unrestricted internet access remains separately governed.
A shared AI service on the school LAN.
Teacher governed
Students can ask questions, practise concepts and search approved learning material through a local interface. The school can decide who uses it, which resources are connected and when internet research is allowed.
A useful learning assistant without handing over the whole internet.
Family controlled
Suitable home hardware can host an approved AI assistant for homework and explanations while parents keep browsing, social platforms and wider internet access under separate household rules.
Selected school and family data can stay on equipment you control.
Purpose limited
Local processing can reduce unnecessary transfer of prompts and documents to external services. Access control, backups and data-retention rules still matter.
Local AI is not a magic child-safety filter.
Human oversight
Models can still be wrong or inappropriate. Schools and parents still need age-appropriate configuration, supervision, acceptable-use rules, digital literacy and safeguarding.
For organisations
Private intelligence for teams that cannot depend on the cloud for everything.
NGOs, lodges, farms, workshops, schools and businesses can run an internal assistant over approved documents and workflows, keep selected information on premises and continue routine work through connectivity interruptions.
Ask approved policies, manuals and procedures.
Local knowledge
Build an internal assistant over operational documents, maintenance manuals, training material or institutional knowledge with access rules appropriate to the organisation.
Use the cheapest or safest intelligence layer for each job.
Flexible
Routine and private workloads can stay local while approved complex tasks use frontier cloud models when their additional capability justifies external processing.
Keep working when the internet does not.
Site ready
A lodge, farm, field office or workshop can keep local manuals, procedures and AI assistance available across the LAN even if upstream connectivity is temporarily unavailable.
Choose who can use what.
Role based
Local deployment can be combined with accounts, role separation, network controls, logging where justified and clear model-update responsibility.
Examples, not promises
The model decides the hardware.
Small models: some current quantized models are compact enough for phones and ordinary computers.
Mid-size models: examples such as Gemma 3 12B can fit an 8 GB laptop GPU in quantized form, while larger 20B–27B-class models often benefit from 16–24 GB or more usable memory.
Large local reasoning models: OpenAI's gpt-oss-20b is designed around roughly 16 GB of memory; gpt-oss-120b targets about 80 GB. Exact requirements still change with context length, precision, concurrent users and runtime.
What Talunza does
We build the deployment around the real workflow.
The expensive mistake is buying a GPU first and asking what to do with it later. We size the workload before recommending hardware.
Use case, users, privacy and performance.
Scope first
We determine what the AI must actually do, how many people will use it, what data should stay local and whether cloud, local or hybrid operation is the better fit.
Model, runtime, access and local network setup.
Quoted
We install suitable software, configure the selected model, connect approved local knowledge where required, and expose a simple controlled interface.
Local AI can become part of a real system.
Custom
Over time it can connect to documents, maintenance records, school resources, field tools, business systems, sensors or Talunza applications where the workflow justifies it.
Updates, security, backups and model changes.
Optional support
Open model weights may be free, but hardware, electricity, storage, cooling, networking and maintenance are not. We make those responsibilities explicit.