The archive · Developer & Business Tools · Product decision · 2022
min-dalle: a solo PyTorch port of DALL·E Mini that reached 3.5k GitHub stars
A solo port stripped DALL·E Mini to inference-only PyTorch with four dependencies; the 2022-06-28 Show HN drew 663 points and the repo 3.5k stars.
min-dalle (kuprel/min-dalle)
What the business is
min-dalle is a free open-source Python package, installed with pip install min-dalle, that runs DALL·E Mini's Mega text-to-image weights as pure PyTorch inference on CUDA or CPU, with numpy, requests, pillow and torch as its only third-party dependencies.
How it started
On 2022-06-28 the developer kuprel posted Show HN: I stripped DALL·E Mini to its bare essentials and converted it to Torch, pointing to min(DALL·E), a fast, minimal port of Boris Dayma's DALL·E Mini with mega weights that had been stripped down for inference and converted to PyTorch, keeping numpy, requests, pillow and torch as the only third-party dependencies.
What happened
The Show HN drew 663 points and 148 comments. The README advertised a 3x3 grid of DALL·E Mega images in 55 seconds on a Colab T4, 33 seconds on a P100 and 15 seconds on an A10G on Hugging Face; commenters benchmarked it on Apple-silicon MacBooks and RTX 3090s, filed setup issues, and the author answered with fixes in the thread. Third parties built on it, including a YouTube walk-through by The AI Epiphany and the NeoGen Discord bot, whose A100 benchmark graph is credited in the README.
How it ended up
The port kept an audience as a free developer tool rather than becoming a company: the repository snapshot taken for this batch shows 3.5k stars, 247 forks and 377 commits, and the README still presents the same minimal, pip-installable port with no business model or paid tier.
Background
min-dalle is a free open-source port by the developer kuprel that re-implements Boris Dayma's DALL·E Mini as fast, minimal PyTorch inference. The model's Flax weights are converted to torch and everything not needed to generate an image is stripped away, leaving numpy, requests, pillow and torch as the only third-party dependencies, installable with pip install min-dalle.
kuprel launched it on Hacker News on 2022-06-28 as Show HN: I stripped DALL·E Mini to its bare essentials and converted it to Torch. The README quoted a 3x3 grid of DALL·E Mega images in 55 seconds on a Colab T4, 33 seconds on a P100 and 15 seconds on an A10G, and commenters benchmarked Apple-silicon MacBooks and RTX 3090s in the thread; the post drew 663 points and 148 comments.
The bet was that developers would run DALL·E Mini-class generation themselves if the port were lean enough for a single GPU or a CPU. As of 2026-09-05 the repository remains live with 3.5k stars, 247 forks and 377 commits, and the project stayed a free tool of the 2022 DALL·E Mini wave rather than becoming a business.
What has to be true
- DALL·E Mini's original JAX/Flax stack carried training-era machinery that had nothing to do with generating one image, so stripping the model for inference removed the real barrier to self-hosting.
- Only numpy, requests, pillow and torch as dependencies made pip install min-dalle viable in environments where reproducing the original model's environment was painful.
- CUDA and CPU paths with float16 and bfloat16 options made the mega model practical on a Colab T4 or even a laptop, not just a datacenter GPU.
- The 2022-06-28 Show HN drew 663 points and 148 comments, and the repository stood at 3.5k stars, 247 forks and 377 commits in the snapshot used here.
What can be applied
A viral model's bottleneck is often runtime, not quality: strip DALL·E Mini to inference-only PyTorch with four dependencies and developers will run it on hardware they already own.
Aftermath
min-dalle remained a project, not a company: as of 2026-09-05 it is still a free, pip-installable open-source package whose repository shows 3.5k stars, 247 forks and 377 commits, a YouTube walk-through by The AI Epiphany, and third-party credits such as the NeoGen Discord bot's A100 benchmark graph. The page presents no paid tier or business model; its role was to give developers a lightweight local route into the DALL·E Mini boom of mid-2022, and that role is what the repository still serves.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card