← 全部模型

GLM · 近 7 天

GLM 5.3

0–100 分,越高代表社区使用体验越正面,不代表能力测试成绩。

更新于 2026-09-05 14:21 UTC · 2026-08-29 – 2026-09-05 UTC
社区口碑分67.5/100136 条评论 · 近 7 天

社区评论

精选近 7 天的不同意见,摘录条数不代表真实好评比例。

26 条精选摘录

正面通用文本

“直到我把它插入GLM-5.3进行故障排除时才发现。感谢这个脚本,它运行得很好。”

Redditen机器翻译
展开原文

Only found out when I plugged it in GLM-5.3 to troubleshoot. Thanks for the script, it works great.

正面安全与拒答

“用非 RE 内容预热上下文配合美国模型 > 切换到 GLM 5.3 处理实际请求,让它用一些 RE 工作填充上下文”

Hacker Newsen机器翻译
展开原文

warm up context with non-RE things with American models > switch to GLM 5.3 with actual request, let it fill context with some RE work

负面安全与拒答

“与其他AI相比,在这次更新和GLM-5.3推出后,它的安全协议非常激进。无论我让她做什么或创建什么,她都会给出一个完全扭曲的回应,说这很危险,可能会造成风险”

Redditen机器翻译
展开原文

Compared to other AIs, after this update and the arrival of GLM-5.3, it has a very aggressive security protocol. Anything I ask her to do or create, she gives off a completely distorted view, saying that it's dangerous and could cause risk

正面本地部署

“我实际上可以同时运行16个GLM 5.3代理。”

Redditen机器翻译
展开原文

I can actually have 16 GLM 5.3 agents running concurrently at the same time.

正面通用文本

“对我来说物有所值,我买了他们的Max计划,用glm 5.3编排规划,小型实现和部署都非常迅速,另外我还额外购买了claude max仅用于fable评审。该死,就这么跑”

Redditen机器翻译
展开原文

Worth the money for me, I got their Max plan, orchestrate planning with glm 5.3, small and implementing things are flash, also I got extra claude max for fable only for review. Shit just run

正面推理

“GLM 5.3现在对我来说真的是GOAT(史上最佳),我有一个大型项目,GPT和Claude都彻底搞砸了,而我刚刚用300百万token的声明和YOLOd zcode在这个项目上搞定了。”

Redditen机器翻译
展开原文

GLM 5.3 is truly the GOAT for me right now, I had a large project that both GPT and Claude completely shat the bed in, and I just got the 300mil tokens claim and YOLOd zcode on the project.

中性速度与延迟

“运行glm 5.3 dsv4 flash等,你可以并行至少4个代理”

Redditen机器翻译
展开原文

Run glm 5.3 dsv4 flash etc you can parallel like 4 agents at least

正面本地部署

“现在就使用glm 5.3,完美契合”

Redditen机器翻译
展开原文

just use glm 5.3 now which perfectly fits

正面通用文本

“用GLM 5.3一枪就搞定了”

Redditen机器翻译
展开原文

one shotted this with GLM 5.3

正面编程

“GLM 5.3在开发工作上与SOL对我来说是同一水平。有时候甚至更好。我用SOL来做路线图/任务,然后用GLM审计并做实际工作。”

Redditen机器翻译
展开原文

GLM 5.3 on same level in dev work as SOL for me. Sometimes even better. I'm using SOL for roadmap/tasks, then GLM audit on them and real work.

正面编程

“在Max推理上使用完整版GLM 5.3(非Flash)在我(编程)使用中大致介于GPT 5.6 Terra-High和Sol-High之间。 我让它日常使用,用Sol来检查计划。这样帮我大大延长了用量。”

Redditen机器翻译
展开原文

Full GLM 5.3 (not Flash) on Max reasoning falls somewhere between GPT 5.6 Terra-High and Sol-High in my (coding) use. I let it daily-drive and use Sol to check plans. Helps usage stretch far for me.

正面编程

“GLM完成工程,kimi做验证并且返回迭代方案”

中性通用文本

“我确实觉得模型在输出中这么频繁地引用时间是不必要的,但它的表达确实很流畅,所以这其实并不是一个负面特性”

Redditen机器翻译
展开原文

I do find it unnecessary for the model to referencing time so much in its outputs but it does flow in my experience so it hasnt really been a negative quirk

褒贬兼有推理

“智商和高级模型坐一桌,它干不了的,大多数情况下glm-5.3也不行”

负面通用文本

“比如 GLM-5.3 的真实泛化能力与海外顶尖梯队仍有差距,但 AA 榜总分却紧咬 GPT-6(只差一分)”

负面推理

“五分钟前它轻松彻底破解了一个复杂的GLM 5.3诊断题,GLM不得不承认失败 lol”

Redditen机器翻译
展开原文

it had no problem completely busting a comprehensive GLM 5.3 diagnosis for me 5 minutes ago to which GLM had to admit defeat lol

负面本地部署

“用GLM 5.3,你的配置会比基准测试显示的效果更差,因为你必须运行比基准测试使用的更低量化版本。”

Redditen机器翻译
展开原文

With GLM 5.3, your setup would do worse than what you see in benchmarks, since you'd have to run a lower quant than what the benchmarks use.

负面本地部署

“你只能容纳Q4量化的GLM 5.3,所以我不认为把它换成DeepSeek的全精度有什么意义。”

Redditen机器翻译
展开原文

You can only fit GLM 5.3 in Q4, so I doubt it makes sense to trade that for full precision on DeepSeek.

负面推理

“知识丰富:我让它默写了一首茉莉花的音乐,做出来八九不离十。同样的任务交给 GLM-5.3,只会重复第一句,后面都想不起来了。”

褒贬兼有角色扮演/创作

“5.1这个群像总是玩着玩着后面就变成机器人讲话了”

负面编程

“几何和多边形很弱”

Redditen机器翻译
展开原文

geometry and polygons were weak

负面通用文本

“两者的楼梯都有客观上很糟糕的问题。”

Redditen机器翻译
展开原文

Both have objectively terrible staircases.

负面编程

“5.3 Flash很棒,实际上比更大的5.3感觉更像下一代产品。”

Redditen机器翻译
展开原文

5.3 Flash is awesome, actually feels like the next generation compared to the bigger 5.3.

正面本地部署

“坚持使用能装进内存的模型(及其上下文窗口),Q4量化的GLM 5.3或Q8量化的GLM 5.3 Flash可能是更好的选择”

Redditen机器翻译
展开原文

stick to models that can fit (+ their context windows) on RAM, Q4 GLM 5.3 or Q8 GLM 5.3 Flash might be better fits

中性编程

“完全换回了GLM 5.3,但把所有视觉需求都转给Flash处理。”

Redditen机器翻译
展开原文

Switched back to GLM 5.3 exclusively but route any vision needs to Flash.

中性本地部署

“Flash需要192gb显存才能流畅推理,而5.3在nvfp4下也能装进同样的192gb(但很紧张)”

Redditen机器翻译
展开原文

Flash needs 192gb vram for nice inference and 5.3 fits same 192gb at nvfp4 (but tight)