A GitHub repository shared on Hacker News demonstrates the DeepSeek V3 architecture implemented in approximately 190 lines of pure JAX code, according to a report by another outlet.reported
What was released or claimed
The project claims to reproduce the model architecture without high-level libraries like Flax or PyTorch, using plain dictionaries for parameters and readable functions for every component.
Who it affects
Machine learning researchers interested in minimal implementations of mixture-of-experts models and developers working with Apple Silicon hardware constraints can reference this code.
What is still unknown
It remains unclear whether this minimal implementation scales effectively to larger datasets or if the performance gains hold outside the specific public-domain fiction training set.
So what
The demonstration suggests that complex architectures like DeepSeek V3 can be distilled into compact, readable code for educational purposes and lightweight experimentation.reported
What the paper shows
The research tests whether the DeepSeek V3 technical report's architecture functions correctly in a constrained environment, reporting results from training on 21.7 million tokens from 207 books.reported
Why it matters
- Researchers can audit the exact mechanics of multi-latent attention and expert routing without library abstraction layers.
- The code proves that bf16 compute with f32 master weights runs on consumer-grade Apple Silicon with 24 GB memory.
- Validation loss remained identical whether expert balancing was enabled, supporting the paper's claim about bias efficiency.
- The implementation avoids specific JAX-metal failures by design, offering a workaround for current hardware driver limitations.
Topics: DeepSeek, JAX, mixture-of-experts, Apple Silicon, open-source AI
Updated
Then / now / next
DeepSeek V3 was primarily accessed through large-scale deployments and official technical reports.
A minimal 190-line JAX implementation allows local training on consumer hardware for educational analysis.
Researchers may test whether this compact structure supports larger datasets or different domain texts.
Story timeline
- Then DeepSeek V3 was primarily accessed through large-scale deployments and official technical reports.
- 14 Nov 2026 Show HN: DeepSeek V3 in ~200 lines of pure Jax
- Now A minimal 190-line JAX implementation allows local training on consumer hardware for educational analysis.
- Next Researchers may test whether this compact structure supports larger datasets or different domain texts.
Try it
The fog came in off the
Good result: The model generated a continuation describing fog, trees and stars, demonstrating basic next-token prediction capability.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Does this model require PyTorch to run?
No, the implementation uses pure JAX and explicitly avoids PyTorch, Flax and Equinox dependencies.
What hardware is needed to replicate the run?
You need an Apple Silicon Mac with 24 GB of memory, Python 3.12 and the uv package manager.
Is the checkpoint safe to load?
The checkpoint is a Python pickle that can execute code, so users should only load it from trusted sources.
Connected stories
Reading receipt
- The implementation uses 190 lines of pure JAX without Flax, Equinox or PyTorch dependencies.
- Training occurred on a MacBook Air M5 with 24 GB of memory using public-domain detective fiction.
- Expert balancing via bias achieved even routing without changing validation loss at step 2000.
Still open: How does the model perform on general-purpose text compared to the specialized crime fiction training data?
Still open: Can the 190-line implementation scale to larger parameter counts without significant code complexity increases?
