Practical AI Guides & a Free Claude Code Course

A company called LiquidAI makes tiny AI models. They say this is their first “omni-model” — meaning, a single model that takes text, image, audio, and video inputs, and produces text or speech outputs. Previously, they had separate models for text and images; now they’ve merged these into a single model. And they’re processing both the audio and visual components of a video simultaneously, rather than separately. This model is tiny — only ~2.2 billion parameters. Parameters are the internal weights learned by a model during training, so lower parameter count leads to smaller and faster models. They designed this model specifically for mobile, laptop, etc. To play around with it, go to this Hugging Face space (a website for hosting AI models & demos): huggingface.co/spaces/LiquidAI/LFM3-Omni-2B-WebGPU. You’ll need a recent browser such as Chrome or Edge. It should load directly in your browser using WebGPU, which allows you to run AI models on your GPU via the browser. There’s nothing to download or install. Ask it questions about pictures (by typing) or videos (by speaking) and see what responses it gives you in text or speech form. It supports over twenty languages!

It’s a very detailed guide named ‘The Ultimate MoE Training Handbook.’ Written by three researchers at Hugging Face (a platform for hosting AI models and tools) and one of the authors is a co-founder of Hugging Face. MoE refers to Mixture-of-Experts, a technique for building large models using smaller ‘experts’, which are only activated for subsets of inputs rather than all inputs. According to Artificial Analysis (a website ranking AI models), 18/20 top open models are based on this method, including DeepSeek, GLM, MiniMax, etc. In this handbook, they explain how to train a MoE model step by step, such as choosing hardware, pre-processing data, training a model with 6.86 billion parameters but only activating 1.22 billion parameters for each input, etc. They also introduce the ‘router’, which selects experts for different inputs, and the load balance among different experts. There are lots of interactive visualizations based on actual training results on NVIDIA H100 GPUs (high-performance GPUs for training AI models). All you need to do is to visit the webpage and start reading, no installation required. But be warned that it will take you a day or two to finish reading!

This is a model card on Hugging Face, a site where people share AI models. The model is called Qwen3.8-Flash-Next-E384-FP4-V2, and it's built on top of a Qwen model, an 80-billion-parameter AI model made by Alibaba. A Hugging Face account named autotrust took that base model and retrained it with reinforcement learning, a method where the model gets rewarded for good answers, to do one job: spot a security flaw in code and write it up as a structured report scored by CVSS, the industry scale for how serious a vulnerability is. It's an unofficial fan-made version, not approved by Alibaba or NVIDIA. To install it, you can't — it's a server model, tested on one NVIDIA B200, a data-center graphics card, not something you run at home. Here you can only read the card and see what changed.

zhanghao123vip/ai-manju-skills

AI Manju Skills is a set of 21 text templates for making "manju" — short serialized vertical video dramas, a popular format in China — one template per genre: urban revenge, palace intrigue, cultivation fantasy, wuxia (martial-arts epics), post-apocalypse, campus romance and more. Each genre is a single file called SKILL.md, a skill file you feed to an AI agent — there's nothing to install. You copy the genre folder you want, for example skills/dushi-manju for urban drama, into ~/.claude/skills/, the folder where Claude Code (Anthropic's coding agent) looks for skills. Then you send the agent one sentence with your story idea, and it runs the whole production pipeline on its own: a project card and plan (story, characters, world, episode-by-episode plot, hooks for future episodes), a prompt library for character, prop and scene images, episode scripts broken into scenes with full dialogue, a director's plan with shot and video prompts for every episode, and finally a release package with a cover-image prompt and announcement text. By default that's 30 episodes, about 4 minutes each, in vertical 9:16 format — you can ask for other numbers. The instructions file is Chinese-only, no English version. The authors say: one episode per chat, one template per chat, or genres and settings start bleeding into each other.

WoYun520/pixel-gif-sticker-pack

This is a skill for Codex (OpenAI's coding agent). You give it one picture of a character and ask for a sticker pack. It builds animated GIFs — short looping picture files — from that one image. Three looks: pixel (small blocky figures, 320x320), clean illustration (anime or chibi characters, 512x512), and photo-real, which keeps a real person's face from the photo. Two motion modes: "light motion" keeps one face and pose with a small wobble, good for churning out a big batch fast; "full action" has the character do a whole gesture — wave, heart-hands, bow — moving head, body and arms together. The repo ships 242 ready examples plus a bakery ad clip that strings mascot moves and paper-cutout props into a tiny story. You get separate GIFs, a zip of the whole set, and a preview; full-action sets also come with a frame-by-frame check sheet. To install: tell Codex "install this sticker Skill: https://github.com/WoYun520/pixel-gif-sticker-pack", then send it a photo and ask for 10 light-motion GIFs with "hi", "thanks", "bye". No app or GPU needed — Codex's model does the work; running the Python scripts yourself just needs Python 3 and Pillow.

Sandflow-ai/sand3d

Sandflow is the team behind sand3d. sand3d is a Python library — Python is a coding language. It lets you build a 3D shape with code, not mouse clicks. The code runs on top of Blender, a free 3D program, through Blender's own Python toolkit called bpy. The makers call it "3D modeling in Python, built for LLMs" — LLMs are AI models like the one behind Claude. A part in sand3d uses 2 to 6 times fewer tokens — tokens are the text chunks an AI reads — than the same part written in Blender's own Python, or in CadQuery or FeatureScript, two other code-based 3D tools. The whole command vocabulary fits in about 3,600 tokens. That's small enough to paste whole into an AI agent's instructions. You can export the finished model to GLB, a 3D file format for games and the web. It also exports to URDF and USD, formats read by robot simulators such as ROS, PyBullet, MuJoCo, and Isaac Sim. The same code always builds the exact same file, bit for bit. To install: grab a normal laptop, no graphics card needed. You just need Python 3.11 or 3.13 with Blender's bpy module — Python 3.12 won't work, and Intel Macs have extra limits. Ask your AI agent to run pip install sand3d, write a program for a simple part following the README's example — README is the project's instructions page — and export it to GLB so you can see the result.