三天烧完 GPT-6 Astra 周额度后转本地跑模型

CyrilXBT @cyrilXBT原推发布于 2026.09.18

在 X 上打开原推

burned through my entire gpt-6 astra weekly quota in three days. replacing my $200/mo codex subscription with something that costs nothing per token, just disk space. it's called colibrì. one developer, pure c, zero dependencies. the actual trick isn't a smaller model, it's refusing to load the whole thing into memory at once. it streams moe experts straight off an nvme drive as it needs them, which is why a 2.8 trillion parameter model runs on a machine that could never hold that much in ram. nine model families supported right now. glm-5.3 flash, kimi k3, deepseek v4 flash, several others. worth knowing before you commit storage, glm-5.3 flash actually needs somewhere around 200 to 230gb of ram for full residency, not the 25gb some posts are quoting, that lower number belonged to the earlier glm-5.2 specifically, and it's easy to conflate the two if you're skimming. kimi k3 doesn't share a codepath with the others at all. it gets its own dedicated engine because the architecture underneath it is genuinely different, hybrid attention combining kimi delta attention layers with gated mla layers, and no positional encoding anywhere in the model, position lives inside the attention mechanism's own convolution and decay instead of being bolted on separately. ``` git clone https://github.com/JustVugg/colibri cd colibri/c ./setup.sh cd .. ./coli doctor ./coli chat --model /path/to/model ``` grew from 4,611 stars in its first eleven days back in july to past 28,000 by mid september. one person built and is still actively maintaining this. the actual tradeoff, stated plainly. no token bill, no weekly quota you can burn through in three days of normal use. what you're paying instead is disk space, ram, and patience, since streaming experts from storage is never going to match a hosted gpu cluster on raw speed. the frontier didn't get smaller. it just moved from your subscription to your ssd. https://x.com/Av1dlive/status/2099816833788567683/video/1

这条是引用着另一条写的：x.com/1373665408193036293/status/210039675107295…