Free AI access
Surprisingly, many people have still never talked to a neural network or run a single AI agent - it seems expensive, complex, or 'not for me'. In fact you can start for free right now, often without a bank card or a powerful computer. Below are proven ways to get access to AI: from free tiers to compute grants. Limits and terms change - we check them regularly.
How to use: pick a card, open the link and sign up - current limits are listed on each one.
Free tiers
Work instantly and for free, usually without a card: sign up and use. There are limits (requests per day, tokens per minute), but they're plenty to get started and for personal tasks.
Gemini 2.5/3 Flash, Gemma 3, multimodal
28-30 free models through one key
GPT-5, GPT-4.1, GPT-4o, o4-mini, Llama
Trial credits
A starting balance in dollars to try out (e.g. $5). Expires in a few days or weeks - handy to test a model for your task before paying.
New Google Cloud accounts; spend on Gemini 3.1 Pro, Imagen, Veo, full Vertex AI catalog
Startup programs
Large compute grants and credits for companies: from thousands to hundreds of thousands of dollars. How to get it: register a startup (legal entity, product, site) and apply; some programs require an accelerator or an investor.
For students
Free access and credits with student status. How to get it: confirm you're a student - a .edu email or verification via GitHub Education and university programs.
Verified students
Free credits for students attending YC campus events 2025-2026; includes Azure/AWS cloud, GPT/Claude/Grok, voice, search, video, databases
Free A100 compute credits with verified .edu email for students; 3–14 day activation
Local AI: your own model on your own hardware
Open models can run right on your computer - free, offline, and your data never leaves it. Below, in plain words, we figure out what your machine can handle and pick a model you install with a single command.
How to check your hardware2 minutes and you know your class.
An AI model is a large file that loads entirely into your computer's memory to run. The smarter the model, the bigger it is and the more memory it needs. Simple rule: the number in a model's name (e.g. 7B = 7 billion parameters) is roughly how many gigabytes of memory it needs; in the compact Q4 format that's about half (7B ≈ 4-5 GB). A model runs fastest on a graphics card, in its video memory (VRAM); without a strong GPU it runs on regular RAM - slower, but it works. Macs with M1-M4 chips use shared memory, so they handle even large models well. First, find out how much memory you have:
- Mac: Open the Apple menu (logo in the top-left of the screen) and choose About This Mac. The Memory line shows your amount - for example, '16 GB'.
- Windows: Press `Ctrl+Shift+Esc` together, open the Performance tab and pick Memory on the left. The total is shown at the top - for example, '16.0 GB'.
- Linux: In a terminal, type `free -h` and press Enter. On the `Mem` line, the first number is your total RAM.
- NVIDIA: In a terminal, type `nvidia-smi` - the VRAM size is in the top-right of the table. Or check it in the GeForce Experience app.
- AMD: On Linux use `radeontop`; on Windows check the AMD Adrenalin app.
- Windows: Press `Ctrl+Shift+Esc` -> Performance -> GPU. Find the 'Dedicated GPU memory' line - this works for any card.
- Mac (M1-M4): There is no separate VRAM - it uses the computer's shared memory, which is a plus: almost all of it can go to AI.
Take a model whose size (in Q4, shown on the card) is smaller than your free memory, and leave 2-4 GB spare - for the system and the 'conversation memory' (context). Not sure? Start with a smaller model: if it runs snappy, try a bigger one. We've already sorted the models into tiers below - just pick your memory size.
Entry-level laptops, base MacBook Air M1/M2, PCs without a GPU.
Best pick for weak hardware: fast, keeps a conversation well and follows instructions.
Strong at reasoning and code for its size, commercial-friendly (MIT).
The lightest: runs almost anywhere, good for simple tasks and notes.
How to run it
LM Studio - a GUI app: install it, pick a model from the list, click download and chat. Easiest to start, no command line.
Ollama - command-line: a single `ollama run <model>` downloads and runs the model. Great for developers and integrations (a local API on localhost).
Model comparison
These are just examples for local use. Browse and compare every model in the full catalog: All models → · Compare models →
| Model | Tier | Size (Q4) |
|---|---|---|
| Llama 3.2 3B★ | 8 GB RAM | ~2 GB |
| Phi-4 Mini | 8 GB RAM | ~2.5 GB |
| Gemma 2 2B | 8 GB RAM | ~1.6 GB |
| Llama 3.1 8B★ | 16 GB RAM | ~4.9 GB |
| Qwen2.5 7B | 16 GB RAM | ~4.7 GB |
| Mistral NeMo 12B | 16 GB RAM | ~7 GB |
| Qwen2.5 Coder 32B★ | 32 GB RAM / VRAM | ~20 GB |
| Gemma 2 27B | 32 GB RAM / VRAM | ~16 GB |
| DeepSeek R1 32B | 32 GB RAM / VRAM | ~20 GB |
| Llama 3.3 70B★ | 64 GB+ RAM / VRAM | ~43 GB |
| Qwen2.5 72B | 64 GB+ RAM / VRAM | ~47 GB |
| DeepSeek R1 70B | 64 GB+ RAM / VRAM | ~43 GB |
FAQ
What is quantization and Q4?
Quantization shrinks a model by lowering weight precision - the file gets several times smaller and fits in memory. Q4 (4-bit) is a popular balance: near-full quality at about half the size of Q8. Sizes in this guide are for Q4.
RAM vs VRAM - which matters?
On a GPU the model runs fastest - here VRAM size matters. Without a strong GPU the model loads into regular RAM and runs on the CPU (slower, but works). Apple Silicon has unified memory, so Macs handle large models well.
Can I use these commercially?
Check the model license. Apache 2.0 and MIT allow commercial use with almost no limits. Llama and Gemma licenses also allow it but with conditions (e.g. a user-count cap for Llama). Always read the license text.
How much disk space do I need?
Each model is a separate file from ~1.5 GB (light) to 40+ GB (large). Leave spare room for several models. Ollama and LM Studio keep them in their own folder; you can delete unused ones with one command.