三天烧完 Astra 周额度，改用本地开源模型跑 GLM/Kimi/DeepSeek
作者称 GPT-6 Astra 周配额三天耗尽，于是用 3.2 万星的 Colibri 在本地无 GPU 跑 GLM-5.3、Kimi K3、DeepSeek V4。
I BURNED THROUGH MY ENTIRE GPT-6 ASTRA WEEKLY QUOTA IN 3 DAYS. SO I’M REPLACING MY $200/MO CODEX SUBSCRIPTION WITH THIS. a free open-source repo with 32,960+ GitHub stars that lets you run models like: → GLM-5.3 Flash → Kimi K3 → DeepSeek V4 Flash locally. here’s the setup: 1. install Colibri download the release for Linux, macOS, or Windows. no compiler needed. 2. or build it yourself git clone http://github.com/JustVugg/colibri cd colibri/c ./setup.sh cd .. 3. choose your model GLM-5.3 Flash → ~195 GB DeepSeek V4 Flash → ~167 GB Kimi K3 → ~1.6 TB the wild part? you don’t need a GPU. 4. get the model weights GLM has its own converter, while DeepSeek and Kimi can load from their original Hugging Face snapshots. 5. run it ./coli chat ./coli web ./coli serve then use: ./coli doctor ./coli plan ./coli tune to check your setup and optimize performance. the tradeoff is pretty simple: no token bill. no hosted quota. no dependency on an API. but you better have the storage. the frontier just moved from your subscription to your SSD.
这条是引用着另一条写的：x.com/mikenevermiss/status/2098270564016091458