🧭 Which AI should I use?
📰 This week in AI
Loading this week's updates…
💡 Tips by tool
✍️ Prompting basics
📊 Which AI for what
| Job | Best pick | Why |
|---|
🖥️ Run AI Locally
Why local? Your prompts stay on your machine, it works offline, and there's no monthly fee once you own the hardware. LM Studio "can operate entirely offline" once models are downloaded (LM Studio), and Ollama says local runs don't send your prompts or data to ollama.com (Ollama FAQ). Trade-offs: local models are usually less capable than frontier cloud models, you pay up front for RAM or a GPU, and setup takes some tinkering.
📄 Free cheat sheet: Which local AI fits your RAM? One page: model picks for 8, 16, 24–32 and 64 GB+, a tool chooser, and when "local" really is private. Sizes checked on ollama.com on Oct 17, 2026.
Hardware: what you need
| Platform | Guidance | Source |
|---|---|---|
| Mac (Apple Silicon) | The CPU and GPU share one pool of "unified" memory, so your RAM is your model memory. LM Studio needs an M-series Mac and macOS 14+, and recommends 16 GB+ RAM (8 GB Macs work only for small models with short context). The GPU can't use all of it: in Apple's example for M1 Pro and M1 Max MacBook Pros, the GPU can access about 21 GB on a 32 GB Mac and 48 GB on a 64 GB Mac (newer chips may differ). | LM Studio · Apple |
| Windows / Linux PC | On Windows, LM Studio recommends 16 GB+ RAM and at least 4 GB of dedicated GPU memory (VRAM), and x64 CPUs need AVX2 support; on Linux it needs Ubuntu 20.04 or newer. Ollama supports NVIDIA GPUs (compute capability 5.0+), AMD GPUs (ROCm or Vulkan) and Apple Metal. | LM Studio · Ollama hardware |
| Rule of thumb | The download size of a 4-bit model is roughly the memory it needs to load, plus extra for context (the KV cache). For example, Ollama lists gemma4:e4b at 6.6–9.5 GB and gemma4:26b at 16–19 GB. | Ollama library |
Tools
LM Studio
A desktop app for downloading and chatting with models on Mac, Windows and Linux. It can also run a local server that apps and agents connect to as if it were the OpenAI API.
Ollama
Install it, then run ollama run gemma4:e4b in a terminal. It runs a local API at localhost:11434, and ollama launch can start agents such as Hermes Agent on a local model.
llama.cpp
The open-source C/C++ engine behind many local apps. It runs GGUF model files on Apple Silicon, NVIDIA (CUDA), AMD, Vulkan and plain CPU.
MLX / mlx-lm
Apple's machine-learning framework for Apple Silicon. Models converted for it ("MLX" versions on Hugging Face) run natively on a Mac's unified memory, and LM Studio can load them. Run pip install mlx-lm (Python 3.11+), then mlx_lm.chat; mlx_lm.server serves an OpenAI-style API on localhost:8080. Ollama also lists MLX builds, such as gemma4:12b-mlx (7.7 GB).
Model size by hardware tier
| Your memory | What fits (4-bit, roughly) | Examples |
|---|---|---|
| 8 GB | Small models (about 2–4B parameters), with short context | Browse by size in the Ollama library. For example, Gemma 4 E2B is listed at 4.6–7.5 GB (listing). |
| 16 GB | Models of about 7–12B parameters | Most open families (Gemma, Qwen, Mistral, Llama and others) publish models in this range; check the file size before downloading. |
| 32 GB | Models of about 20–30B parameters, or mixture-of-experts models (only a few billion parameters active per token, so faster) | For example, Gemma 4 26B-A4B MLX 4-bit is 15.6 GB (HF). |
| 48–64 GB+ | Models of about 30B+ parameters with longer context | Check the model card's listed size, and leave room for context. |
These are general guides, not endorsements of one model. Check the license, and where the model comes from, before you use it (see the foreign AI safety note below).
Quick setup
Ollama in 2 steps
1) Install from ollama.com. 2) Run ollama run gemma4:e4b and start chatting. Other apps can then use http://localhost:11434.
LM Studio on a Mac
1) Install LM Studio. 2) Search for a model and pick an MLX build (for example "gemma-4-26b-a4b 4bit"). 3) Load it and chat, or start the local server for other apps to use.
Local agents
Hermes Agent is a real, public agent from Nous Research. It runs on Linux, macOS, WSL2 and Windows, works with "any endpoint" (including local servers), and connects to Telegram, Discord, Slack and more (docs, GitHub). Ollama can start it on a local model with ollama launch hermes --model gemma4 (Ollama).
Example: a hybrid setup
A common pattern: a 32 GB-class laptop runs a mid-size MLX model in LM Studio for private everyday work, and an agent falls back to a small local model through llama.cpp when cloud services are unavailable. Speeds vary widely by chip, model and context length.
Local vs cloud cost: local costs nothing per month after the hardware (plus electricity), while cloud plans run from about $5 to $500 a month (see below). Cloud models are generally stronger and need no setup. Many people use both, with local for private or offline work and as a fallback.
🌍 AI from other countries
| Service | Company / base | Notes |
|---|---|---|
| DeepSeek | Hangzhou DeepSeek AI (China) | Its cloud app stores data in the PRC (privacy policy). Open-weight models are available to run locally. See the safety note below. |
| Qwen | Alibaba (China; Qwen Chat is run by Alibaba Cloud Singapore) | Open-weight models such as Qwen3.5-9B. Qwen Chat's policy reportedly says data is processed in Singapore and mainland China (third-party summary of the policy). |
| Kimi | Moonshot AI (China) | Kimi says consumer data is stored in mainland China and de-identified chats may be used for training (Kimi help). It says API data isn't used for training (API FAQ). |
| Mistral (Le Chat) | Mistral AI (France, EU) | Subject to EU privacy rules (GDPR). Training on your data can be switched off (Mistral help). |
🛡️ Safety note: AI services based in China
- Where your data goes: DeepSeek's cloud service says it collects, processes and stores personal data in the People's Republic of China (policy), so Chinese law applies. China's 2023 generative AI rules require services to uphold "core socialist values" (peer-reviewed analysis).
- Censorship: Studies find DeepSeek R1 refuses some China-sensitive political questions, even when run locally (R1dacted, arXiv).
- Past exposure: In January 2025, Wiz found a publicly accessible DeepSeek database containing over a million log lines, including chat history and API keys. DeepSeek secured it promptly after being told (Wiz).
- Government restrictions: Australia ordered DeepSeek removed from all government systems and devices (PSPF Direction 001-2025). Italy's data protection authority restricted DeepSeek's processing of Italian users' data (Garante). Several US agencies and states restricted it on official devices (TechCrunch).
- Cloud app vs local: When you run open weights offline (LM Studio, Ollama, llama.cpp), your prompts never reach the developer's servers. Any built-in censorship still remains, and you should download models only from official or well-known repositories.
- Bottom line: These models can be excellent and are often free. Don't paste sensitive personal or work data into any cloud chatbot you don't trust, and follow your employer's rules.
💳 Paid AI plans compared (US)
| Service | Tiers & US price / month | Key limits & notes | Source |
|---|---|---|---|
| ChatGPT | Go $8 · Plus $20 · Pro $100 / $200 / $500 | Monthly billing only, with no annual option. Pro 500 includes "Ultrafast". Since Oct 29, Pro 100 and Pro 500 include twice their previous Work and Codex allowance, and the planned Pro 200 cut was canceled (all Pro 200 plans now have the higher allowance). Pro 200 is open to new subscribers again. | chatgpt.com/pricing · Pro tiers |
| Claude | Pro $20 ($17 billed annually) · Max $100 · Ultra $200 | Usage resets on a rolling 5-hour window, and paid plans also have weekly limits. Pro gets at least 5x Free's usage; Max gets 5x Pro's every 5 hours and every week; Ultra gets 10x Pro's weekly usage with no 5-hour limit for verified accounts. Max 5x and Max 20x were renamed Max and Ultra on Oct 30. Includes Claude Code. | claude.com/pricing |
| Gemini (Google AI) | Plus $4.99 (2 TB option $9.99 on Google One) · Pro $19.99 · Ultra $99.99 (5x) / $199.99 (20x) | Includes Google storage: 400 GB (Plus; 2 TB on the $9.99 Plus option), 5 TB (Pro), 20 TB or 30 TB (Ultra 5x / 20x). Ultra gives 5x or 20x Pro's usage limits. | gemini.google · Google One |
| Grok (The X Subscription) | Plus $8 · Premium $30 · Super $100 · Ultra $200 (annual pricing available) | A weekly usage pool shared across Grok and Cursor: Premium 4x, Super 15x and Ultra 50x Plus's usage. Existing SuperGrok plans continue. | x.ai/pricing · docs |
| Perplexity | Pro $17/mo billed annually · Max $200 ($167/mo billed annually) | Higher Research, file and model limits on Max. Max prices from Perplexity's Max help page. | perplexity.ai/hub/pricing |
All prices re-checked November 9, 2026. US prices before tax; they change often, so confirm on the official page before buying. No affiliate links.
📬 Get the free weekly digest
Five practical, sourced AI tips every Tuesday. No hype. Free, unsubscribe anytime.
Also free, no signup needed: our one-page cheat sheet Which local AI fits your RAM? (PDF).
👋 About AI Tips Weekly
AI Tips Weekly is a free, five-minute email for people who use AI at work and want it to actually save time. Each Tuesday you get five practical tips for tools like ChatGPT, Claude, Gemini, Gemini Notebook (formerly NotebookLM) and local models. Every tip includes how to do it, its limits, and a link to the official source so you can check it yourself.
What you won't get: hype, recycled listicles, or recommendations we can't back up. If we ever link to a tool we earn from, it's labeled right next to the link.
⚖️ Disclosure & affiliate policy
Disclosure: Some links on AI Tips Weekly may be affiliate links. If you buy through them, AI Tips may earn a commission at no extra cost to you. We only include tools we think fit the tip, and the commission never changes what we recommend. (There are currently no affiliate links or sponsors on this page.)
- Fit first. We only link to a tool when it directly helps with the tip.
- Disclosure near the link. Affiliate links are labeled right beside them, not only in a footer.
- Plain words. "Affiliate link", "I earn a commission" or "Sponsored". No vague tags.
- Honest. We note downsides and free alternatives. No invented results, reviews or scarcity.
- Sponsors are separate. Labeled, visually separate, and with no editorial control.
- Re-verified. Program terms and claims are re-checked, and links to closed programs are removed.
Based on the FTC Endorsement Guides: FTC guidance · 16 CFR Part 255