# Vaibhaav-Tiwari/deepseek-v3-jax > A from-scratch implementation of the DeepSeek-V3 architecture (MLA, DeepSeekMoE, aux-loss-free balancing, MTP) in 190 lines of pure JAX, trained on a MacBook Air on public-domain detective fiction. It is an educational architecture demo, not a strong language model. - Magnitude: 5.5 out of 10 — Early signal - Stars: 210 total · +12 stars measured 2026-11-16, 00:28–09:36 UTC, ≈ 38 by evening - Star trust: star growth looks organic - Category: Language models · Language: Python · License: MIT · Created: 2026-11-13 · Last push: 2026-11-15 - GitHub: https://github.com/Vaibhaav-Tiwari/deepseek-v3-jax · Homepage: https://vaibhaav.in/writing/deepseek-v3-jax/ · Page: https://gitnova.dev/en/r/Vaibhaav-Tiwari/deepseek-v3-jax ## Useful for - Study the DeepSeek-V3 architecture by reading the 190-line implementation top to bottom - Reproduce the training run on an Apple Silicon Mac and compare validation loss and expert load - Run the balancing ablation with and without the bias to verify the paper's claim ## Why it’s here - Star-counter measurements on 2026-11-16 (UTC), 00:28–09:36: 198 → 210 stars (+12). This is the change over that interval. - Estimated end-of-day forecast: about +38 stars, using observed gains and the previous day. - The repository is 3 days old and already has 210 stars. With less than two weeks of history, there's no usual pace to compare the spike against yet. - Top new repositories this week: #164. - Hacker News: “Show HN: DeepSeek V3 in ~200 lines of pure Jax” — 273 points, 1 day ago. - About 9 forks a day — people are taking the code. - Recent forks include notable developers: @jonnydubowsky (307 followers). ## Star trust Star growth looks organic. Star-trust labels are heuristics based on the repository’s behavior, not a check of every stargazer. ## Numbers - Forks: 18 - Issues and pull requests: 1 - Watchers: 0 - Average over the last week: 62 per day - Usual pace: too little history (under two weeks) - Stars in the last hour (measured): 1 - Latest release: v1.0 (2026-11-14) ## Stars per day, last 4 days (oldest → newest, today is partial) 2026-11-13 … 2026-11-16: 0, 154, 55, 12 ## Magnitude by day, last 3 days 11-14 5.4, 11-15 6.9, 11-16 5.6 ## Hacker News - Show HN: DeepSeek V3 in ~200 lines of pure Jax — 273 points, 28 comments: https://news.ycombinator.com/item?id=50497589 ## Spotted in now - Top new repositories this week: #164 - Spotted on Hacker News ## More in this category 1. **ggml-org/ggml** — 7.3 · Breakout · Language models · C++ · +134 stars measured 2026-11-16, 00:28–09:36 UTC, ≈ 290 by evening A dependency-free C/C++ tensor library for machine learning with quantization and support for CPU, GPU, NPU, and WebAssembly. It serves as a foundation for running ML models across platforms with minimal setup. Full card: https://gitnova.dev/en/r/ggml-org/ggml.md 2. **basetenlabs/deepseek-ultrafast** — 5.0 · Early signal · Language models · Rust · +14 stars measured 2026-11-16, 00:28–09:36 UTC, ≈ 42 by evening A high-performance Rust inference engine for DeepSeek-V4.1-Flash on sm_100a/sm_103a GPUs, supporting 1M context and multi-server KV cache sharing. Full card: https://gitnova.dev/en/r/basetenlabs/deepseek-ultrafast.md 3. **hyperionaidev/smolmaxxing** — 4.3 · Early signal · Language models · C · +11 stars measured 2026-11-16, 00:28–09:36 UTC, ≈ 32 by evening A project for running LLMs on low-power devices (TV boxes, old tablets, SBCs) with ARM-specific optimizations: custom kernels, weight streaming, MoE expert paging. The goal is to squeeze maximum tokens per second out of devices with 1-2… Full card: https://gitnova.dev/en/r/hyperionaidev/smolmaxxing.md 4. **Helldez/BigMoeOnEdge** — 4.2 · Early signal · Language models · C++ · +43 stars measured 2026-11-16, 00:31–09:36 UTC, ≈ 96 by evening An engine for running large MoE GGUF models on devices with insufficient memory: experts are streamed from disk as needed, built on llama.cpp. Works on Android, iOS, macOS, Windows and Linux, including CPU, Metal, CUDA, HIP, Vulkan,… Full card: https://gitnova.dev/en/r/Helldez/BigMoeOnEdge.md 5. **ggml-org/llama.cpp** — 4.0 · Steady · Language models · C++ · +76 stars measured 2026-11-16, 00:29–09:37 UTC, ≈ 183 by evening LLM and VLM inference implemented in C/C++ with no dependencies, running on CPUs and GPUs across many hardware backends. Enables local model execution with minimal setup and quantization. Full card: https://gitnova.dev/en/r/ggml-org/llama.cpp.md --- Magnitude (0–10) measures how fast and how unusually interest in a repository is growing right now. It is not a quality score. Days are UTC. “So far today” is a fact; “expected by the end of the day” is a forecast. Summaries and use cases are written by an LLM (DeepSeek V4.1 Flash) from the README and may be inaccurate: verify specific claims (benchmarks, speed, hardware) in the repository itself. Data as of 2026-11-16 09:47 UTC, updated every 30 minutes.