AINewsnow

189 lines, 39 tests pass: I ran DeepSeek-V3 in pure JAX

This story is from 2026-11-15. It is preserved in the archive; the latest stories are on the live feed.

189 lines, 39 tests pass: I ran DeepSeek-V3 in pure JAX A Show HN post this week claims the whole DeepSeek-V3 architecture — MLA, DeepSeekMoE, aux-loss-free balancing, multi-token prediction — in 190 lines of pure JAX, pretrained from scratch on a MacBook Air. I counted 189 lines of actual code, ra…

Read the full story at DEV Community — Machine Learning ↗

Timeline · 1 report

  1. 2026-11-15 15:17 · DEV Community — Machine Learning
    189 lines, 39 tests pass: I ran DeepSeek-V3 in pure JAX

More stories

  1. Sarvam Launches India-Hosted Token Factory, Faces Backlash Over Chinese Model Offerings — Bloomberg AI
  2. Tencent launches Hy4 Flash at $0.07 per million tokens as AI price war intensifies — Business Insider AI
  3. AI Models Regularly Identify Civilians as Military Targets in Simulated Warfare Tests — Bloomberg AI
  4. DeepSeek Removes Separate Fees for Model-Triggered Search and Web Fetch — Pandaily
  5. Run Mobilerun with local and cheap models Ollama, DeepSeek, Qwen, OpenRouter — DEV Community — Machine Learning
  6. Open-weight coding models had a big October. How a solo builder should decide whether to switch. — DEV Community — AI
  7. Same AI model, five weeks later: the cheapest provider had changed — DEV Community — AI
  8. DeepSeek Desktop: balance, reasoning and task controls in one workspace — DEV Community — AI

Read the Daily Brief: the day's most important AI stories →