low level engineering @massive_com | creating @fusor_rs | building cool stuff — ex @supabase
Orlando, FL
Joined June 2020
- Tweets1.3K
- Following162
- Followers592
- Likes3.5K
Pinned Tweet
qwen3.5 2b reading a receipt at 84 tok/s. in chrome, on an m4 pro cpu.
no server. no gpu. zero network requests after load.
follow-ups start in 0.14 s. rust compiled to wasm with fusor-ai 🦀
new model in the browser every week. next up: indra ocr.
huggingface.co/spaces/fusorl…
drop a 4 second clip into a chrome tab, ask two questions about it, and both answers come back in 3.2 s. the clip never leaves the page.
i ported d3-lite from vllm's semantic router with fusor-ai. it never writes text, it just scores the options.
Huggingface: huggingface.co/spaces/fusorl…
this is fusor's own website running as an iphone and ipad app, unchanged.
Hyperpane turns any fusor app (html + rust) into an iphone, ipad or mac app with one command, without writing any swift.
v0.1, ios is simulator-only for now.
github.com/fusor-rs/hyperpan…
It's weird hearing your own voice say something you never said.
Clone your voice, locally, running in the browser, with WASM.
Record a few seconds of yourself, type anything, and it talks back as you. qwen3-tts on the cpu, built with fusor-ai. the recording stays on your machine.
Demo: huggingface.co/spaces/fusorl…
Source: github.com/fusor-rs/qwen3-tt…
LFM3 has been everywhere this week, so i wanted to see if it would run in a browser tab.
it does. all six models, on the cpu, built on fusor-ai (wasm) runtime 🦀. the text one writes 97 tokens a second.
still early, so if you try it, let me know what breaks
huggingface.co/spaces/fusorl…
source below 👇
Source for WASM website: github.com/fusor-rs/liquidai…
Fusor AI: github.com/fusor-rs/fusor-ai
qwen3.5 2b reading a receipt at 84 tok/s. in chrome, on an m4 pro cpu.
no server. no gpu. zero network requests after load.
follow-ups start in 0.14 s. rust compiled to wasm with fusor-ai 🦀
new model in the browser every week. next up: indra ocr.
huggingface.co/spaces/fusorl…
Andres Pirela retweeted
For the first time in the history of AI models, a model has survived an entire month without losing its crown.
Opus 5.5 needs to be studied.
Looks promising!
Andres Pirela retweeted
Samsung should be criminally prosecuted if this is true.
live-tweeting on samsung findings were fun.
i'll let my agents give it a bit more time, but this document is already the most comprehensive breakdown of some key concerns of tizen i've seen run on a live device.
it takes like a good while for me just to scroll through it (there's about 44 writeups as of right now on individual functionality).
release is probably happening in a few days.
Qwen 3.5 2B running fully in the browser. No API calls. No hops. Just fully Qwen in Web Assembly.
Should I release it? 👁️
I think I just found the fastest OCR model I've ever tested.
Running OCR entirely in the browser with WebAssembly and Fusor AI (@fusor_rs).
Same benchmark. Wildly different results:
🐢 PaddleOCR (@PaddlePaddle): 20 seconds
⚡ OXR (@somarkai): 5 seconds
⚡ ⚡⚡⚡⚡⚡Indra OCR (@datalabto): 0.12 seconds
That's ~42x faster than OXR and ~167x faster than PaddleOCR.
All running on WASM.
This is insane.
This picture was generated fully on the browser, with WASM. 44 seconds across 8 threads.
Follow what we're doing at @fusor_rs
I love how for Anthropic this was a simple release, for OpenAi it took hype, DevDay, vague posting, all to end up in the same place: being behind.
WHAT THE ACTUAL FUCK?!
Anthropic just decided to go after OpenAI Dots and Grok Bot with Claude Duet, powered by the absolute MONSTER that is Opus 5.5.
We literally go to sleep, wake up, and somehow Anthropic has dropped ANOTHER product to take on its competitors.
LIKE HOW ARE THESE GUYS SHIPPING THIS FAST?!
THEY JUST CAN'T STOP WINNING.
What an absolutely LEGENDARY run. 😭
Andres Pirela retweeted
Two small open TTS models dropped this week: Ant Group's Ming-3.0-tts-0.5B and Gradium's Phonon-31M.
Within a day, @apirela_dev had both running locally in the browser on CPU. He used fusor-ai, a Rust toolkit he built that compiles models into WebAssembly programs.
His benchmarks on an M4 Pro:
- Phonon — 2.4x faster than Gradium's browser CPU runtime with 8 threads, with a 48.5 MB download vs. 69.4 MB.
- Ming — first audio in 0.3s vs. 14.3s for the non-streaming PyTorch reference, with quantized weights.
He also shares where his implementation loses: Gradium's native CPU runtime is 3.5x faster on a single thread.
"I'm all for privacy models running on CPU," he says.
Both implementations are open source and still at v0.1.
phonon-wasm.fusor.build
ming-wasm.fusor.build
#BackchannelSignals
This quoted post is unavailable.
Introducing Parakeet WASM: speech to text that runs entirely in your browser
116 s of audio -> 3 s. On your CPU. No server. No GPU. No API key. Your audio never leaves the tab.
Moondream's Parakeet Redux, built with fusor-ai for WASM compatability.
Try it now: parakeet-stt.fusor.build
it transcribes at real-time speed, and does not use GPU. It's great to embed in edge hardware, or frontends without having to pay a subscription.
This was made with love by @fusor_rs for @moondreamai
<3