AI Tutorials for Companies and Teams
Many coders love working directly in the command line. We prefer VS Code for almost everything, because you get code, files, terminal, and AI plugins cleanly in one place.
Agent-first IDE from Google. Runs as a native development environment on your machine, not in the browser.
Currently the best price-performance ratio for many. Students often get it free, and you can compare multiple models directly in the editor.
Strong in chat for structure, explanations, and longer coding tasks.
Best via the VS Code plugin if you want to try Qwen directly in the editor.
Our recommended setup. Codex, GitHub Copilot, Claude, and Qwen run here in one place with files, Git, and terminal.
If OpenAI, use Codex instead of ChatGPT for coding. Strong for agentic work directly in the tool and editor.
Describe the goal, user, platform, and desired result as concretely as possible.
Currently, it's best to prompt and code in English. German also works but usually a bit worse. Most models are primarily trained on English, and the programming world's language is English.
Start with a small project rather than a huge idea.
First, let the model build a simple file and folder structure.
Ask the model to briefly explain each step and only make small changes at a time.
Small landing page
To-do app
Notes app
Simple data or CSV viewer
First let it plan, then implement step by step.
Briefly test after each step.
Copy errors directly back into the tool and ask for a concrete fix.
llama.cpp, Ollama, or LM Studio let you bring local LLMs into VS Code or your own setups.
Git or GitHub for version control
A terminal for commands
A good editor like VS Code
A clear README with goal and setup
Tauri Framework
Native Swift Framework
Xcode Project
Everything runs locally on your machine. Slower than the AI providers, but very secure and costs only electricity, no subscriptions. Almost all local LLM programs are based on
llama.cpp, an important open-source library.
Very simple llama.cpp wrapper. A bit like ChatGPT-style for local models.
Slightly more complex, but still simple wrapper with more configuration options.
For pros who want to integrate local LLMs into their own tools or work directly with the command line.
Database for current and distilled models.
Smaller, trimmed versions of large models. They're faster, cheaper, and often good enough for many practical tasks.
Simple: 3B to 7B, often ~2 to 5 GB
Medium: 8B to 14B, often ~5 to 10 GB
Advanced: 32B and up, often 20 GB+
Rule of thumb:
16 GB RAM or VRAM is the absolute minimum.
More memory usually means larger models, more context, and smoother work.
Computer with at least 16 GB RAM
or a strong Nvidia GPU
The more memory, the better.
🔢 4B, 6B, 8B etc. roughly indicate the number of parameters in a model.
More parameters usually means more knowledge, nuance, and complexity. But it also requires more memory and compute.
🪟 GGUF is the standard format for many local
llama.cpp setups and usually the best choice on Windows.
🍎 MLX is especially interesting for Apple Silicon and local models on Mac.
Other models we use locally for coding: Qwen, Ministral, and GLM.
Qwen and SmolVLM.
Currently, Macs are the best and simplest setup for local coding with AI, mainly because of Unified Memory: CPU, GPU, and Neural Engine share the same memory pool. This often makes local models on Mac more straightforward and efficient.
Mac Studio with at least 32 GB memory.
If you're developing with Unreal Engine, for PlayStation, or generally doing Windows-heavy work, a Windows machine is the better choice.
MacBook Pro (Apple M Chip) with 16 GB RAM
Mac Studio (Apple M Chip) with 32 GB RAM
Windows PC with RTX 4070 Super and 64 GB RAM
Claws are personal AI assistants or agent runtimes that don't just chat. They take on real tasks: web, apps, files, terminal, and entire small workflows.
The most important open-source Claw right now. Originally started by Peter Steinberger. Local, flexible, and clear in the agent and assistant space.
Cloud version around OpenClaw. One-click setup, 24/7 online, and significantly easier for beginners.
Open-source agent runtime by Nous Research with persistent memory, its own CLI, an API server, and connections to more than 20 chat platforms.
OpenAI's direction for real computer-use and web tasks. Can browse, use files, fill out forms, and execute actions.
Anthropic's computer-use approach. Can see screens and directly control mouse and keyboard.
Established providers are releasing their own personal assistants or programs that can directly control the computer. The direction is clear: less chat, more real execution.
OpenAI: ChatGPT agent, formerly Operator
Anthropic: Claude Computer Use
Google: Project Mariner
Kimi: Kimi Claw, Kimi Agent Swarm
These tools often get extensive access: browser, files, terminal, logins, clipboard, or entire apps.
Prompt injections are a real problem. A website or file can slip hidden instructions to the agent.
If you run something like this with broad permissions, in the worst case you're voluntarily installing a trojan.
Therefore: only test on machines with little sensitive data, grant few permissions, treat skills and plugins like real code, and never blindly confirm everything.
ZeroClaw, NanoClaw, PicoClaw, TrustClaw, Nanobot, Grip AI, AutoGPT, CrewAI, Dify, n8n, Open WebUI, LibreChat, LobeChat, Jan, AnythingLLM, Manus, Perplexity Computer, Claude Cowork, Vellum
A quick overview of which models we're currently tracking for chat and coding. This is an assessment, not a complete market overview.
