Check your device
Use the Tinbot Device Checker first. Write down RAM, free storage, GPU/VRAM, operating system, and what job you want the device to do.
Open the checker
The Local Intelligence Project
The Local Intelligence Project
A practical first path for people who checked their device and want the next step: install a runner, download a model, test local chat, then add documents or agent tools one layer at a time.
This page is the public preview of the TLIP Local AI Starter Pack. The GitHub version is the maintained source for exact model links, issue tracking, community notes, templates, and update dates.
Our first starter pack for running local AI is live on GitHub: setup guides, model links, downloadable templates, and knowledge-pack examples. Check back for updates as the guides become community-tested.
Use the Tinbot Device Checker first. Write down RAM, free storage, GPU/VRAM, operating system, and what job you want the device to do.
Open the checkerUse LM Studio for a beginner GUI, Ollama for a simple local service/API, or llama.cpp when you want the lower-level engine and GGUF control.
See runner linksStart smaller than your machine can theoretically handle. A small model that runs smoothly teaches more than a large one that freezes.
See model tiersAsk a few normal questions, then test speed, heat, memory pressure, and whether the answer quality is useful enough for your task.
Coming soonUse a local knowledge base or RAG tool before fine-tuning. For most people, documents plus search is safer, faster, and easier to correct.
Coming soonOnly after local chat works should you add memory, agents, tools, personas, or fine-tuning. Each layer needs its own clear guide.
Coming soonA runner is the software that loads the model weights on your computer and gives you a way to use them. It is the bridge between the downloaded model and the chat window, command line, or local API.
Exact requirements depend on quantization, context length, runner, GPU support, and the model itself. The GitHub repo should keep the dated model matrix.
Best first test for old laptops, low-memory devices, phones, and learning machines.
Practical floor: 4-8 GB RAM, 10+ GB free storage.Good first everyday local chat tier when a device has enough memory and storage.
Practical floor: 8 GB RAM, 20+ GB free storage.The common local-AI target for useful chat, coding help, and small document workflows.
Practical floor: 16 GB RAM, SSD preferred, GPU helpful.Better reasoning and writing, but heavier memory pressure and slower CPU-only use.
Practical floor: 24-32 GB RAM or suitable GPU/Apple unified memory.Useful for shared branch machines, but not a normal beginner setup.
Practical floor: 64 GB+ memory class, strong GPU or high-memory workstation.This is a starter index for model families people commonly test locally. Exact model choices, licenses, and hardware notes will live in the GitHub model matrix.
Common local path: small Llama family models for first tests, larger 70B+ models for strong shared machines.
Start: 1B-8B tierGood beginner family to track because it spans small everyday-device models through stronger single-GPU classes.
Start: 1B-4B tierOpen-weight reasoning models from OpenAI. The 20B model is the practical local target; the 120B model belongs on high-memory GPU systems.
Start: 20B tierBroad family with chat, coder, embedding, vision-language, and reasoning variants. Keep runner and license notes visible.
Start: 0.6B-8B tierUseful 7B, 12B, 24B, and mixture-of-experts options. Better for tested machines than first-install guesswork.
Start: 7B-12B tierTrack small distilled reasoning models separately from very large MoE releases; hardware needs vary sharply.
Start: 1.5B-8B tierMIT-licensed flagship model for long-horizon coding and agentic work, with a 1M-token context window. Its 744B-total, 40B-active MoE checkpoint is about 1.51 TB in BF16, so this is a multi-GPU cluster or hosted-API model, not a normal laptop download.
Start: research cluster / APIImportant open-weight lineage, but not a normal beginner download. Grok releases are large research or branch-host scale models, with Grok-2 documented as a multi-GPU, hundreds-of-GB checkpoint.
Start: branch host / researchLightweight models are important for reclaimed laptops, edge devices, and teaching the workflow before scaling up.
Start: 2B-4B tierA directory for models beyond this starter list. People and organizations publish model repositories with weights, files, model cards, licenses, and community variants; verify the publisher, license, and hardware requirements before downloading.
Directory: more model families