DeepSeek V3 reimplemented in 200 lines of Jax
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Developer Vaibhaav Tiwari released deepseek-v3-jax, a 190-line pure JAX rebuild of DeepSeek V3's architecture.
It shows the model's key components can be coded compactly without PyTorch or Flax.
Background
DeepSeek V3 is a large language model developed by DeepSeek, and its technical report describes the architecture that Vaibhaav Tiwari, the author of the deepseek-v3-jax repository, set out to reproduce. The repository implements that architecture in 190 lines of pure JAX — a numerical computing library — with no Flax, Equinox or PyTorch, and was pretrained from scratch on a MacBook Air on public-domain detective fiction. It drew attention after a Show HN post on Nov 14, 2026, and a companion write-up with diagrams, code explanation and results.
How it unfolded
- Nov 14, 2026A Show HN post titled "DeepSeek V3 in ~200 lines of pure Jax" is published.
- Nov 14, 2026Reports say Vaibhaav Tiwari released deepseek-v3-jax, a 190-line pure JAX implementation covering features including multi-head latent attention and DeepSeekMoE.
- recentlyThe deepseek-v3-jax repository is public with 4 commits, 3 stars and 1 fork.
Key terms
- JAX
- A Google-developed Python library for high-performance numerical computing and large-scale machine learning.
- PyTorch
- An open-source deep learning library for building and training neural networks, originally developed by Meta.
Source: Wikipedia
Who’s saying what
- Author
- The repository README says the DeepSeek-V3 technical report architecture is implemented in 190 lines of pure JAX, pretrained from scratch on a MacBook Air on public-domain detective fiction, with no Flax, Equinox or PyTorch.