Edge0-145B runs on one 16 GB NVIDIA card, and Brave sells search to AI agents

On Friday's Hacker News front page, Edge0 released Edge0-145B, an open-weight model built from Qwen3.8-Flash-Next to run on your own computer. It has 145 billion parameters, but each token draws on only about 10 billion of them, and on one RTX 5090 it writes about 80 tokens a second, faster than you can read. On agent work in a terminal it edges past its parent model and past Claude Opus 4.6, and it keeps working offline, with no bill per token.

Edge0-145B needs less than half the memory of the model it came from. Each group of 64 experts shares two large matrices, and each expert keeps only a small set of weights of its own, so far less has to be stored and moved. The experts used most stay in graphics memory, the rest wait in system memory, and vocabulary data comes off the SSD when it is needed. That is how a card with 16 gigabytes, plus 20 gigabytes of system memory, runs the whole model.

Edge0 promised macOS on day one. The build that shipped runs on Linux and Windows, with an NVIDIA card and CUDA 13. In the thread, coder543 noticed that the team benchmarked an Apple M2 Ultra and still left the Mac runtime out of the release. That Mac number measures an unreleased build. On a Mac today, the smaller Edge0-35B already runs on Metal.

The Brave Search API now takes payment through x402, an open protocol built on the HTTP status 402, Payment Required. An agent asks for a search, gets back a 402 with the price, signs a payment in USDC on the Base blockchain, and asks again. Each search costs half a cent and carries its own payment, so an agent with a funded wallet needs no Brave account and no API key. In the thread, ddxv wrote that they put x402 on a heavily crawled site and have yet to see a single payment come through.

scottjg wanted a Coway Airmega air purifier on the Good Night button in Apple Home, without Coway's cloud in the middle. During setup, the purifier downloads Amazon's root certificate over plain HTTP, so scottjg swapped in their own certificate and could read every connection it made. An AI agent then worked out the Bluetooth setup from the Android app in about 20 minutes, and a new Home Assistant integration sets the purifier up from a browser and talks to it directly. In the thread, Aurornis says models now carry people through the tedious parts, which is why weak choices like this keep turning up.

GET http://rootca-fs.iocareus.com.s3.ap-northeast-2
    .amazonaws.com/AmazonRootCA1.pem

A model's published weights can reveal which NVIDIA chip created them. PyTorch fills a new model's starting weights from a seeded random generator, split across 2,048 threads for every core on the GPU, so the chip's core count decides where each number lands. Weights for tokens the model never trains on keep those starting values. Matched against them, six models from DeepSeek, Moonshot AI, MiniMax and Ant Group, all released since April, fit Blackwell's core counts, while earlier models from the same labs fit the older Hopper chips. The pattern can't show where the chips were. Blackwell servers can be rented in data centers outside China, and in the thread, flowerlad argues that doing so breaks no current export rule.