Watch on YouTubeWatch on Instagram
About this video
A startup called Edge0 says its new open-weight AI, Edge0-145B, has 145 billion parameters and runs on your own computer, offline. Here's what it claims, and the fine print.
Would you run it at home, or keep paying for the cloud?
- For scale: OpenAI's GPT-3 had 175 billion parameters (2020). Apple's on-device model has about 3 billion (Apple, 2024).
- Edge0's own numbers: 29 GB peak memory and 80+ tokens a second; only about 10 billion of the 145 billion parameters work on each step (a mixture of experts), with the rest waiting in RAM and on the SSD.
- Edge0's own tests: 86.5 vs 78.2 for Claude Opus 4.6, an older Claude model, on Terminal-Bench 2.1 (coding agents). Edge0 ran every model's score itself; no independent test has been published yet.
- The fine print: its Hugging Face model card lists Alibaba's Qwen3.8-Flash-Next as the base model. On a PC, Edge0's release notes ask for a gaming graphics card with at least 16 GB plus 20 GB of RAM. The Mac version (Apple Silicon, macOS 26+) arrived Saturday, two days after launch, and Edge0's only Mac speed test used an M2 Ultra, a desktop chip.
- The licence: free for personal use, research and using what it writes, but selling the model or running it as a paid service needs a separate licence from Edge0. Hacker News commenters note that makes it open weights rather than fully open source.
Sources:
- Edge0 blog, "Introducing Edge0-145B" (12 Nov): https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer
- Hugging Face model card (base model, speed table, licence): https://huggingface.co/Edge0/Edge0-145B-NVFP4
- GitHub release v0.2.0 (Windows and macOS packages, requirements): https://github.com/Edge0-AI/Edge0/releases/tag/v0.2.0
- Hacker News discussion: https://news.ycombinator.com/item?id=50475713
- GPT-3: Brown et al., "Language Models are Few-Shot Learners" (OpenAI, 2020): https://arxiv.org/abs/2005.14165
- Apple's on-device model: https://machinelearning.apple.com/research/introducing-apple-foundation-models
Pictures:
- Laptop animation, offline chat demo and the 3D Manhattan scene: Edge0's own launch and demo videos (from its Hugging Face page), shown for news reporting and labelled as Edge0's.
- Edge0's architecture slide and logo: Edge0 (its blog), shown for news reporting.
- Screenshots: Hugging Face (model card, licence), GitHub (release notes), Hacker News (thread), captured 14 Nov.
Music: Deliberate Thought by Kevin MacLeod, CC BY 3.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Deliberate_Thought_(ISRC_USUAN1100261).mp3)
Follow @PranavInnovates for the day's tech news.
What the video says
A startup says its new AI has a hundred and forty-five billion parameters, and runs on your own computer. For scale: OpenAI's GPT-3 had a hundred and seventy-five billion. Apple's on-device model: about three. It's from Edge0, out this week. In its demo, it even runs with the Wi-Fi off. Edge0 says it peaks at twenty-nine gigabytes of memory, and types faster than you read. The trick: only about ten billion of them work on each step. On its own tests, it beats an older Claude model at coding-agent tasks, eighty-six to seventy-eight. The fine print: its model card says it's built on Alibaba's Qwen. On a PC, it needs a sixteen-gigabyte gaming graphics card. The Mac version landed Saturday, and Edge0's only Mac speed test used a desktop chip. And it's not fully open source: a paid service built on it needs Edge0's permission. Developers on Hacker News want independent tests first. Would you run it at home, or keep paying for the cloud?