Demo Check
Claude API · bring your own key
Run a Claude demo 5 to 20 times before you show it once: pass rate, latency, every failure checked by code, and a recorded passing run to fall back on.
Lab
New models come out every day. This is where I try them before building them into anything: what they can and can't do, and how they run on your own device. Nothing to install, nothing uploaded.
By the month each model came out, newest first. The ones that run here say how big they are and where they will run before anything downloads, and stay on your device after the first visit. The ones too big for a browser link to a demo instead.
Built on Claude
Study tools
Study Lab
ML · Audio · Circuits · VR · Design · Physics · Biology · Anatomy · Muscles · Elements
Ten hands on subjects: machine learning from regression to LLMs and agents, sound and audio, live circuits, how VR works, UI design and UX laws, physics simulations, biology for pre-med, full body anatomy in 3D, muscles on a 3D body and the periodic table. Runs in your browser.
Strands Decider 2B
Strands Labs · Apache 2.0
A decision model: it does not write, it picks. Give it a situation and some options and it gives each one a calibrated probability in a single pass.
1.1 GB
ZUNA1.1
Zyphra · Apache 2.0
A foundation model for EEG, the electrical signals of the brain. It can rebuild a missing electrode from the others and turn messy recordings into something models can learn from.
Too big to run here: 380 million parameters, about 1.5 GB.
Matrix-Game 3.0
Skywork · Apache 2.0
A world model: instead of a game engine, a neural network imagines the next frame of a 3D world as you move through it with the keyboard and mouse.
Too big to run here: more than 5 billion parameters and a large GPU.
Voxtral Mini 4B Realtime
Mistral AI · Apache 2.0
Speech to text that keeps up as you talk, in many languages. A WebGPU version exists, but it is a multi-gigabyte download, so here it is a demo instead.
Too big to run here: 4.4 billion parameters, several gigabytes.
FLUX.2 [klein] 4B
Black Forest Labs · Apache 2.0
A small, fast image generator and editor from the makers of FLUX. Only the 4B size is Apache licensed; the larger one is not.
Too big to run here: 4 billion parameters, about 8 GB.
ACE-Step 1.5
ACE Studio and StepFun · MIT
Write a prompt and lyrics, get a full song with vocals in seconds on a good GPU.
Too big to run here: several gigabytes of weights.
Soprano
Eugene Kwek · Apache 2.0
Type a sentence and hear it read aloud, in a clear and natural voice. It is small enough to run in a browser tab and starts talking before it has finished the sentence.
115 MB to 206 MB
AIFS 2.0
ECMWF · CC BY 4.0
The machine-learning weather model from the European Centre for Medium-Range Weather Forecasts, released with open weights.
Too big to run here: about 1 GB of weights and global weather data as input.
TRELLIS.2
Microsoft · MIT
One picture in, a textured 3D model out, ready for a game engine or a VR scene. The weights are MIT; the full pipeline also loads two helper models with their own licences.
Too big to run here: 4 billion parameters plus helper models, many gigabytes.
DPDFNet
Ceva · Apache 2.0
Takes the noise out of speech: cafe chatter, fans, rain. It is small and streaming, built for earbuds and calls, and here it cleans a whole clip several times faster than it takes to play.
10 MB
Teach it by example
Google · Apache 2.0
Google's Teachable Machine idea from 2017, rebuilt on a modern image model: show it a few examples of two things, and it tells them apart live. Nothing is trained on a server; it all happens in your browser.
I found the idea in December 2025. Released Feb 2025.
55 MB to 186 MB
FourCastNet 3
NVIDIA · Apache 2.0
A neural weather model that forecasts the whole planet days to weeks ahead in minutes on a single GPU, a job that takes physics models a supercomputer.
Too big to run here: gigabytes of weights and global weather data as input.
X-VLA
Tsinghua AIR · Apache 2.0
One model that drives many different robots, from single arms to two-armed setups, by learning a small "prompt" per robot body.
Too big to run here: about 0.9 billion parameters plus a robot to drive.
Chronos-2
Amazon · Apache 2.0
Give it any series of numbers, steps, sales, power use, and it says what probably comes next, with a band for how sure it is. It has never seen your data and needs no training on it.
479 MB
G1 Walking Policy
Julien Blanchon and mjswan · Apache 2.0
A humanoid robot that taught itself to walk in simulation, running live in your browser. Shove it and watch a small network catch its balance, fifty times a second.
20 MB
gpt-oss-20b
OpenAI · Apache 2.0
OpenAI's first open-weight language model since GPT-2. It reasons step by step, uses tools, and runs on a laptop with 16 GB of memory, just not inside a browser tab.
Too big to run here: 21 billion parameters, about 13 GB.
Qwen-Image-Edit
Alibaba Qwen · Apache 2.0
Edit a photo by describing the change: swap the background, fix the words on a sign, turn a character to face the other way.
Too big to run here: 20 billion parameters, tens of gigabytes.
mdbr-leaf-ir
MongoDB · Apache 2.0
A tiny search model that matches by meaning, not by keywords. Here it searches everything I have built and written, right in your browser.
24 MB
RF-DETR Nano
Roboflow · Apache 2.0
Point a camera and it boxes and names what it sees, many times a second. A detection transformer small enough for a phone, with no clean-up pass for overlapping boxes.
25 MB to 29 MB
Boltz-2
MIT Jameel Clinic and Recursion · MIT
Predicts the 3D shape of proteins and how strongly a drug-like molecule binds to them, an open alternative to AlphaFold 3 for that job.
Too big to run here: hundreds of millions of parameters and a big GPU.
SmolVLA
Hugging Face · Apache 2.0
A small vision-language-action model: it looks through a robot's cameras, reads an instruction like "put the cube in the box", and outputs the arm's next moves. Trained on community robot data and small enough for a gaming laptop.
Too big to run here: 450 million parameters, and it needs a robot or a simulator to drive.
EdgeTAM
Meta · Apache 2.0
Meta's on-device take on Segment Anything 2. Tap anything in a photo and it lifts out as a clean cut-out, ready to drop into a game or a VR scene.
21 MB to 41 MB
ME-rPPG
Tsinghua University · MIT
Your pulse, from your webcam. Each heartbeat shifts the colour of your skin very slightly, and this tiny model turns those shifts into a pulse wave and a heart rate. The video never leaves your device.
10 MB
Evo 2
Arc Institute · Apache 2.0
A language model for DNA across all of life. It can predict which mutations break a gene and even write new genetic sequences.
Too big to run here: 1 to 40 billion parameters.
Depth Anything V2
HKU and TikTok · Apache 2.0
Give it any photo and it works out how far away every pixel is, so the photo tilts like a small 3D scene. Try it on your phone and move it around.
19 MB to 50 MB