DeepSeek V3 in ~200 lines of pure Jax
Spotted 2026-11-15 05:07 (UTC+8) · original event 2026-11-13 · sources: Hacker News, GitHub · spreading on 1 platformHeat score
Rank score: 6040% — where it sits on the charts
Discussion score: 8040% — how much people are talking about it
Cross-platform score: 3020% — how many platforms it spread to
Composite heat score:42.0rank × 0.4 + discussion × 0.4 + cross-platform × 0.2 — includes a -20-point cooling decay (peak passed, attention fading).
Why it matters
A 671B-parameter frontier model is normally unreadable production code; a 190-line distillation lets ordinary developers actually read how MLA and MoE fit together instead of taking the architecture on trust.The angle
DeepSeek V3 in 190 lines of JAX: what the tiny reimplementation reveals about how the 671B model actually worksSuggested format: Short video · for Developers, AI learners, JAX usersEvidence chain
• HN #10, 101 pts/6 cmt (2026-11-15)• GitHub Vaibhaav-Tiwari/deepseek-v3-jax: DeepSeek V3 (MLA, DeepSeekMoE, aux-loss-free balancing, MTP) in ~190 lines of pure JAX, trained on a MacBook Air, with a write-up at vaibhaav.in/writing/deepseek-v3-jax• repo created 2026-11-13 (GitHub API), 57 stars• kblip/headstart/insighthubmedia listings are HN-mirror aggregators and do not count as platforms (round 54 Terminactory precedent), so cross stays 30• was a 60.0 watch item in round 55• 2026-11-15 06:07 update: HN #10, 118 pts/7 cmt• score unchanged 68.0 (rank 85, discussion 70, cross 30)• 2026-11-15 07:07 update: HN #12, 134 pts/7 cmt• rank 85 -> 75• 2026-11-15 08:07 update: HN #11, 147 pts/8 cmt• score unchanged 64.0 (rank 75, discussion 70, cross 30)• 2026-11-15 09:07 update: HN #9, 166 pts/8 cmt• rank 75 -> 85• 2026-11-15 10:07 update: HN #11, 179 pts/9 cmt• rank 85 -> 75• 2026-11-15 11:07 update: HN #13, 189 pts/10 cmt• score unchanged 64.0 (rank 75, discussion 70, cross 30)• 2026-11-15 12:07 update: HN #14, 200 pts/11 cmt• discussion 70 -> 80• 2026-11-15 13:07 update: HN #19, 207 pts/11 cmt• score unchanged 68.0 (rank 75, discussion 80, cross 30)• 2026-11-15 14:07 update: HN #20, 213 pts/12 cmt• score unchanged 68.0 (rank 75, discussion 80, cross 30)• 2026-11-15 15:07 update: HN #18, 219 pts/20 cmt• score unchanged 68.0 (rank 75, discussion 80, cross 30)• 2026-11-15 16:07 update: HN #20, 225 pts/21 cmt• score unchanged 68.0 (rank 75, discussion 80, cross 30)• 2026-11-15 17:07 update: HN #21, 229 pts/21 cmt• score 68.0->62.0 (rank 75->60, discussion 80, cross 30• below entry line, row kept, score only)• 2026-11-15 18:07 update: HN #23, 228 pts/59 cmt• score unchanged 62.0 (rank 60, discussion 80, cross 30)• 2026-11-15 19:07 update: HN #23, 237 pts/22 cmt• score unchanged 62.0 (rank 60, discussion 80, cross 30)• 2026-11-15 20:07 update: HN #23, 241 pts/23 cmt• score unchanged 62.0 (rank 60, discussion 80, cross 30)• 2026-11-15 21:07 update: HN #24, 246 pts/23 cmt• score unchanged 62.0 (rank 60, discussion 80, cross 30)• 2026-11-15 22:07 update: HN #30, 251 pts/23 cmt• score unchanged 62.0 (rank 60, discussion 80, cross 30)• 2026-11-15 23:07 update: dropped out of HN top 30• cooling -20 (62.0 -> 42.0)