Show HN: DeepSeek V3 in ~200 lines of pure Jax
Hacker News · 15 Nov 2026, 01:04 AEST
DeepSeek V3 has been developed in approximately 190 lines of pure JAX and trained on a MacBook Air. The model features innovations like Multi-head Latent Attention and auxiliary-loss-free balancing.
Key facts
- DeepSeek V3 (MLA, DeepSeekMoE, aux-loss-free balancing, MTP) in 190 lines of pure JAX, trained on a MacBook Air.
- The architecture from the DeepSeek-V3 technical report in 190 lines of pure JAX, pretrained from scratch on public-domain detective fiction.
- Params are plain dicts of arrays and every part of the model is a function you can read top to bottom in deepseek_v3.py.
Read aloud
DeepSeek V3 has been developed in approximately one hundred ninety lines of pure JAX and trained on a MacBook Air. The model features innovations like Multi-head Latent Attention and auxiliary-loss-free balancing. DeepSeek V3 MLA, DeepSeekMoE, aux-loss-free balancing, MTP in one hundred ninety lines of pure JAX, trained on a MacBook Air. The architecture from the DeepSeek-V3 technical report in one hundred ninety lines of pure JAX, pretrained from scratch on public-domain detective fiction. Params are plain dicts of arrays and every part of the model is a function you can read top to bottom in deepseek v3.py.
Sources in this cluster
- Show HN: DeepSeek V3 in ~200 lines of pure Jax
Hacker News · 15 Nov 2026, 01:04 AEST