Endmare

，

[Endmare @3ndmare]

@scottastevenson @willchen500 Not really a secret anymore — Kimi (and other Chinese AI labs) have been hiring top-uni grads to do RLHF for a while now, and through Kimi chat/work, you get plugin access to 177M Chinese court cases + 5.92M statutes. Lawyers in China are actually using this day to day.

[Endmare @3ndmare]

@minchoi meanwhile glm 5.2 is 768b?

[Endmare @3ndmare]

K3 慢的不行啊... 不用子代理，根本用不完199订阅的5小时额度😇

[Endmare @3ndmare]

@1greencat 神奇，用 OS X的竟然还这么多？

[Endmare @3ndmare]

@YuLin807 之前实测过蓝v就55美元等值API，昨天简单用用感觉依旧在这个区间，与之对比Cursor差不多是6倍杠杆

[Endmare @3ndmare]

@JeffreyCalm https://t.co/DH5DAwC4LB?

[Endmare @3ndmare]

@0xkeenz @passluo 试了几个问题，内部知识差太多了，很难是微调

[Endmare @3ndmare]

怎么每个平台一沾上变现内容就开始变味？知乎、B站、小红书、X…… 最狠的算是小红书了，首页发现一堆AI贴、引战帖、进去一看都是卖货的、ai相关话题一堆二道贩子、部分卖家还拉个小群，内部堪比鹅腿阿姨、群体狂热

[Endmare @3ndmare]

@Khazix0918 我也碰见过 4.8 的幻觉问题，对了这个还有工具调用文本标签泄露问题，很难想象这是 2026 S2 的模型，给我干哪来了？

[Endmare @3ndmare]

@sanadi37 真会这样随机地跳舞啊

[Endmare @3ndmare]

@Ansel_Flipradio 在我个人 Agent 用例里都会先进行搜索，这是在什么软件里的结果呢？我感觉智谱已经对这些工具调用进行定向 rl 了，信息不足稳定触发搜索，而不是靠内置知识回答。毕竟它这个模型参数量才 700 多 b，内置知识肯定是很缺乏的。

[Endmare @3ndmare]

@scavenger869 @FireworksAI_HQ GLM-5.2 是自己部署的 fp8精度，里面支持的别的模型也有来自together fireworks groq deepinfra的

[Endmare @3ndmare]

@HarryPa60759488 @9hills Pro 20USD/月 5h/周限形式 感觉比智谱 Pro 用量多 官方是说按 GPU 时间计费的 但是不透明

[Endmare @3ndmare]

@kunoo So sad, So sad...

[Endmare @3ndmare]

@__oQuery 加个 rpiv pi advisor 啥的 ，glm 5.1 语言风格好 gpt 5.5 细心 结合一下效果非常好

[Endmare @3ndmare]

openrouter是公司用有什么deal吗 感觉不如走Claude官方合算？

[OpenRouter @OpenRouter]

Fable is now seeing twice the usage volume of Opus 4.8

(Same daily token usage, but twice the price)

[Endmare @3ndmare]

@byvince_ @0xSero @Snixtp So... should we wait and see if it turns cheap at 5.5 (Fable)?

[Endmare @3ndmare]

@tulipking @NewtypeTrades Bro knows China fr