Fine-tuned Llama-3.1-8B on a 4 GB laptop GPU.
Peak VRAM 3.32 GB. 119.6 tok/s. 100% SM.
The base never fully loads — it streams from RAM one decoder layer at a
time while the card computes the previous one.
Bit-exact vs ordinary training. Paper + code below.
Soup hit 10,000 GitHub stars. we’re gonna need a bigger pot
we’re building open-source tools to train your own models on your own hardware
thank you for cooking with us 🍜
Today we’re launching FlashDrive with @ycombinator 🚀
FlashDrive is a shared cloud drive where your team and all your AI agents work in the same files across computers and cloud sandboxes.
> Search the whole drive with grep, regex or code. Our custom indexing layer finds matches without downloading everything.
> Files stay in the cloud, and your agent only streams the part it needs.
> Start in Claude Code on your Mac. Pick up in Codex.
> Edit the same file at the same time. FlashDrive merges changes or keeps both versions.
> Connect over MCP or mount the drive on macOS or Linux. Give a sandbox access to just one folder for as little as five minutes.
Dalton and I are building the filesystem the next generation of companies will run on. And we're calling it @flashdrive_ai
↓ Try it free! Link below
hey
I’m Maussym, a software engineer from Astana, Kazakhstan
currently building at Soup and leading our global intern team
a little about me:
> 3 years of hands-on development, from websites to full-stack products and AI integrations
> worked across AI startups, education and nonprofit projects
> mentored 50+ learners and now help students build their first hackathon projects
> 3 student research publications
> graduated college with honors, now studying software engineering at Astana IT University
outside of tech: cooking, gym, a little music, and whatever hobby I pick up next ^^
here to share what we’re building at Soup, and what I learn along the way
say hi if you’re building something too :)
hey, i'm alpamys, 18y.o founder from Kazakhstan 🇰🇿
Founder & cto of @trysoup_dev
we're building open-source tools to fine-tune AI models on your own laptop
i'll be sharing what we're building and learning along the way
if you're into AI, open source or building startups, let's connect!
@anhnt256@trysoup_dev Thanks man! Qwen and Mistral are good starting points, but performance depends on your hardware and setup. What GPU are you running?
Soup has raised $2 million!
What started as an open-source project is becoming something much bigger.
Huge thanks to our community and everyone who believed in us. This is just the beginning.
We're growing our team worldwide!
https://t.co/QfDznTbUtO
Speko Week of Voice (Day 2/5)
100+ voice models benchmarked.
Introducing Speko Benchmarks, your source of truth for voice AI models.
Compare speech-to-text, text-to-speech, LLMs and speech-to-speech on accuracy, latency, quality and cost. Explore multilingual performance, code-switching and turn-taking.
New voice model? Check Speko.
Hey there! here’s Maussym, team lead at Soup, on why we’re building Soup Tasters.
Meet the team and see where it takes you 👇
https://t.co/V7kAQovWXe
Build it. Prove it. Own it.
Everyone's talking about Laya, the open 30ms decision model. I spent an evening testing it on one job: replacing the LLM judge in RL fine-tuning, which is the expensive part of every loop.