Edge0

If you’re building toys, glasses, earbuds, or robots, I’d love to hear what you’d put it in. edge0.ai Edge0-AI/edge0

7/ 𝗰𝘂𝗿𝗶𝗼𝘂𝘀 𝘁𝗼 𝘁𝗿𝘆 𝗶𝘁? code: Edge0-AI/Edge0 model: huggingface.co setup help: edge0.ai

𝗧𝗿𝘆 𝗘𝗱𝗴𝗲𝟬 𝗼𝗻 𝗚𝗶𝘁𝗛𝘂𝗯: Edge0-AI/Edge0 𝗠𝗼𝗱𝗲𝗹 𝗶𝗻𝗳𝗼: huggingface.co 𝗝𝗼𝗶𝗻 𝗼𝘂𝗿 𝗗𝗶𝘀𝗰𝗼𝗿𝗱 to get help with setup or share what you’re building: discord.gg 6/6

RUN A 35B AI MODEL LOCALLY ON YOUR PHONE OR PC how to set it up: 1. choose model for hardware - iOS: github.com/Edge0-AI/Edge0/tree/main/ios - macOS: github.com/Edge0-AI/Edge0/tree/main/macos - Android: github.com/Edge0-AI/Edge0/tree/main/android - Windows: github.com/Edge0-AI/Edge0/tree/main/windows 2. follow your guide to build and launch Model info: huggingface.co Setup help / Discord: discord.gg

320,000 GitHub stars in 11 days, should you trust open-source AI? Open-source AI repos are blowing up on GitHub, and one AI tool gained 320,000 stars in 11 days. Here are the craziest open-source AI projects of October 2026, what each one actually does, and whether you should trust open-source AI over closed models like ChatGPT and Claude. We break down REA (Reverse Engineer Anything), the MCP tool that lets an AI agent take apart any app. Then Meta's 816 billion parameter Muse Spark Open, and why Meta still keeps its best model closed. A 35B model running in under 3GB of RAM with Edge0. Fish Audio's Drama 3 voice model. PhotoCraft, a free Photoshop clone rebuilt in Rust. Laya's benchmark fine print. And 722 maths papers from an OpenAI model nobody can use. Then the part most videos skip: what "open source" really means for AI models, open weights vs open source, the 4.5 million fake GitHub stars for sale, China's proposed export controls on AI, and who carries the safety risk when you run a model yourself. Which of these would you actually install? Tell me in the comments. Chapters: 00:00 Reverse Engineer Anything 01:06 Why open source 01:55 Meta draws a line 03:01 Big model, tiny memory 03:46 Voices you can direct 04:29 Photoshop, rebuilt in Rust 05:11 Laya's fine print 05:54 OpenAI's open maths 06:33 Who draws the line 07:35 Pros and cons Sources: REA (GitHub): morluto/rea REA: what is true and what is not: pasqualepillitteri.it Meta Muse Spark Open (PyTorch blog): pytorch.org Meta keeps its top model closed (Briefs): briefs.co Edge0 (GitHub): Edge0-AI/Edge0 Fish Audio Drama 3 (Hugging Face): huggingface.co PhotoCraft (GitHub): storytold/photocraft Laya (GitHub): NandhaKishorM/laya OpenAI math papers (GitHub): openai/math Microsoft + NVIDIA on open-weight risk (Help Net Security): helpnetsecurity.com Open-weight usage and China export controls (The Register): theregister.com 4.5M fake GitHub stars (CyberInsider): cyberinsider.com Open Source AI Definition (OSI): opensource.org Subscribe to @our_consciousview and tap the bell. I check what's real in tech, every episode. #OpenSourceAI #AI #GitHub

10 New GitHub Repos You Should Know This Week (35B AI in 2.5 GB of RAM & More) This week's ten most interesting new open source projects on GitHub: from a tool that removes Apple Intelligence to a 35 billion parameter AI model that runs in about 2.5 GB of RAM. Every pick explained in about a minute, straight from its README. Chapters 0:00 Intro 0:20 #10 omlahore/RemoveMacAI – Get your Mac's space back 0:51 #9 BlueHeisenberg/SCSKiller – Kill shader stutter before you play 1:23 #8 Droid-Deck/DroidDeck – Steam Big Picture on Android 1:54 #7 nykooi1/vibe-wise – Vibe coding that teaches you 2:20 #6 rehan-remade/universal-modder – AI that mods your PC games 2:51 #5 Vincentwei1021/anything2explainer – Type a topic, get an explainer video 3:17 #4 firelex/jeff – A 0.8B model that speeds up agents 3:55 #3 FlashML-org/FreeVideo – AI video on an 8 GB GPU 4:33 #2 storytold/photocraft – Photoshop, rebuilt in Rust 5:06 #1 Edge0-AI/Edge0 – A 35B model in 2.5 GB of RAM 5:42 Outro Repos #10 omlahore/RemoveMacAI: omlahore/RemoveMacAI #9 BlueHeisenberg/SCSKiller: BlueHeisenberg/SCSKiller #8 Droid-Deck/DroidDeck: Droid-Deck/DroidDeck #7 nykooi1/vibe-wise: nykooi1/vibe-wise #6 rehan-remade/universal-modder: rehan-remade/universal-modder #5 Vincentwei1021/anything2explainer: Vincentwei1021/anything2explainer #4 firelex/jeff: firelex/jeff #3 FlashML-org/FreeVideo: FlashML-org/FreeVideo #2 storytold/photocraft: storytold/photocraft #1 Edge0-AI/Edge0: Edge0-AI/Edge0 Scripts written from each project's README; check the repos for details. Music: "Voxel Revolution" by Kevin MacLeod (incompetech.com), licensed under Creative Commons: By Attribution 4.0 License creativecommons.org

Edge0 GitHub Overview: Streaming 35B MoE Models from SSD with 2.9 GB RAM edge0 by edge0-ai: edge0-ai/edge0 Edge0 is an open-source streaming Mixture-of-Experts framework built around a different answer to the memory wall: keep inactive 4-bit expert weights on SSD, predict the next expert with a pre-router, and use Recover-LoRA to restore quality inside a small active-RAM budget. This overview explains how a 35-billion-parameter model is presented as running locally on Apple Silicon Macs and Android phones at roughly 2.9 GB of active memory, while noting the storage-bandwidth, context-length, and mobile-startup tradeoffs. Maintainer-published comparisons cited here report an average 3.9-point gap between the compressed Edge0 pipeline and the uncompressed 35B teacher model. TimeStamps: 0:00 The Memory Wall for Dense 35B Models 0:41 Edge0's Local 2.9 GB Memory Target 1:09 Sparse Experts and 4-Bit SSD Offload 2:04 The Prerouter's Double-Shift Latency Hiding 2:48 Recover-LoRA and Quantization Quality 3:13 Maintainer Benchmarks and the 3.9-Point Gap 3:42 NVMe, UFS, and Long-Context Limits 4:29 Android's Cold-Pool Startup Tax 5:08 What the Architecture Enables—and Where It Bends 🧠 Dense-model memory limits and sparse expert routing 💾 SSD offload, memory mapping, and predictive prefetch 📊 Recover-LoRA quality retention and benchmark tradeoffs 📱 Mobile storage, cold starts, and context constraints The practical takeaway is a systems tradeoff: Edge0 shifts pressure from active memory to fast flash bandwidth and careful routing. The overview is most useful for readers evaluating local MoE inference, Apple Silicon or Android deployment constraints, and the limits of a low-memory headline. #Edge0 #MoE #LocalAI [de-DE] YNAS ynp_5cd9b40fd4d55219 filmisch Privater automatisierter Staging-Upload. Paketidentität: ynp_5cd9b40fd4d55219; Ausgabe: cinematic. [es-US] YNAS ynp_5cd9b40fd4d55219 cinematográfico Carga de entorno de prueba de automatización privada. Identidad del paquete: ynp_5cd9b40fd4d55219; salida: cinemática. [fr-FR] YNAS ynp_5cd9b40fd4d55219 cinématique Téléchargement de l'environnement de test d'automatisation privé. Identifiant du package : ynp_5cd9b40fd4d55219 ; sortie : cinématique. [hi] YNAS ynp_5cd9b40fd4d55219 सिनेमैटिक निजी स्वचालन स्टेजिंग अपलोड। पैकेज पहचान: ynp_5cd9b40fd4d55219; आउटपुट: सिनेमाई। [id] YNAS ynp_5cd9b40fd4d55219 sinematik Unggahan staging otomatisasi pribadi. Identitas paket: ynp_5cd9b40fd4d55219; output: sinematik. [it] YNAS ynp_5cd9b40fd4d55219 cinematic Caricamento di automazione privata. Identità del pacchetto: ynp_5cd9b40fd4d55219; output: cinematico. [pt-BR] YNAS ynp_5cd9b40fd4d55219 cinematográfico Upload de teste de automação privada. Identidade do pacote: ynp_5cd9b40fd4d55219; saída: cinemática.

Edge0 is open source: 35B and 8B models, plus the runtime and inference code. Code: Edge0-AI/edge0 Models: huggingface.co The launch thread from @SamuelZengML: