DeepSeek-V3 architecture reproduced in 190 lines of pure JAX
What happened
On November 14, 2026 at 23:04, a Show HN post appeared on the Hacker News front page about a developer's pure-JAX implementation of the DeepSeek-V3 architecture. The title puts the code at about 200 lines; the post itself gives 190. The implementation uses no Flax, Equinox or PyTorch, and the developer also ran pretraining from scratch on a MacBook Air. The post covers the architecture, code size, dependency choices and the pretraining machine. It gives no training scale, performance numbers or comparison with the original model.
Written by AI from the coverage · updated 2 days ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageShow HN: DeepSeek V3 in ~200 lines of pure Jax
开发者用 190 行纯 JAX 复现了 DeepSeek-V3 架构,不依赖 Flax、Equinox 或 PyTorch,并在 MacBook Air 上从零预训练。
Heat over time
Heat now 3·Comparable peak 9(Nov 15 02:00)·Comparable change over 24 hours -50%
The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.