以下部分引用 AI 总结,现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
GLM
164 atoms · 跨 29 天 · 首见 2026-04-22 · 最近 2026-07-02
三色: 🟦 fact 62 · 🟥 take 102 · stance ▲63/▼26/◆35
叙事状态: 🍂 fading (退潮) · 💤 沉寂 · 策展近14d 0 atom · 🐦 X 近14d 11
状态只算低频策展源 (群/卖方); X firehose 仅作背景音量
来源: X 164
时态: fresh:149 · aging:14 · stale:1
标签: 好观点:81 · 好数字:39 · 好信源:30 · 好思考:23 · 好问题:11
🟨 AI 综合 · junior analyst 概览
展开 AI 综合 (灰色 · 非市场结论 · 点击数字溯源到原 atom)
model: deepseek-chat · 2026-07-12 · 默认折叠
GLM 5.2 的通用能力跃升令人印象深刻,从5.1到5.2的进步远超代码领域,感觉像是突破了某些瓶颈
¹¹。在推理效率上,GLM 仅需其他模型 1/5 的 token,实现相同成本下约 3 倍速度,推理效率排名高于 DeepSeek
¹¹。然而,其 KV Cache 缓存架构存在缺陷,无法卸载到低成本磁盘,导致极端流量下缓存命中率在 5 分钟内骤降至仅 25%
¹¹¹。近期传闻称 GLM 的能力提升可能依赖对 Claude Code 的蒸馏或拦截,尽管蒸馏已不再是驱动力的说法也存在
¹¹。在商业成本方面,GLM 5.2 的 API 成本为 $59.60,虽比 Opus 4.8 便宜 5 倍、比 Fable 5 便宜 11 倍,但自部署需要 2 万美元硬件,且 5.5 年才能回本
¹¹¹¹。值得注意的是,智谱的发布节奏明显加速,GLM 5.1 在 V4 发布前两周就已更强,显示出项目进展加快
¹。目前 GLM 是唯一能胜任我某个复杂代码库挑战的开源模型
¹;但 Jeffries 认为 GLM 5.2 可能创造另一个 DeepSeek 时刻,几乎与 Anthropic 在企业市场平起平坐
¹。
🧭 拥挤度 (一人一票): ▲ 27 位作者 (KOL27) vs ▼ 12 位作者 (KOL12)
⚖️ 多头 27/27 来自KOL
📄 广泛报道的事实 · 2 个
展开 (多人转述同一事实/数字 — 确认度高, **非独立观点**, 不标多空)
- 5.2 发布 — 3 源报道 (含1快讯) ·
X
- 上下文窗口长度 — 2 源报道 ·
X
💢 核心分歧 (4 轴)
1. 模型能力跃迁 vs 持续依赖蒸馏
*topic: 技术路线/竞争格局/监管政策 · 15 bull vs 3 bear*
🟢 bullish 侧:
- <a id="atom-65f184b35e704e4b"></a>2026-06-13
X·@AiBattle_ [老化中] · 5.2 is a significant improvement over 5.1 · 产品发布/技术路线
GLM 5.2 seems to be a significant improvement over GLM 5.1 and has a Max Thinking option
tags: 好观点·好问题 · → daily
📷 原图
- <a id="atom-dbac054c07ed939e"></a>2026-06-17
X·@teortaxesTex · DSa架构扩展到多模态 · 技术路线
no fundamental problems in expanding DSA architectures to multimodality
tags: 好观点 · → daily
📷 原图
- <a id="atom-1472e4556890932f"></a>2026-06-18
X·@thegenioo · 模型需要增加视觉能力和更快的推理速度 · 技术路线/产品发布
The model is insane they just need to give it vision and they need to get inference faster
tags: 好观点 · → daily
- <a id="atom-1acfb4ae9118d50b"></a>2026-06-18
X·@teortaxesTex · 能力差距缩小 · 技术路线 +同日2条
Is reduces the gap more than R1 did at its time.
tags: 好观点 · → daily
📷 原图
- <a id="atom-387d8ed54109cdc3"></a>2026-06-20
X·@teortaxesTex [老化中] · coding stability · 技术路线/模型能力
it's vastly more stable in coding and general agency than any of them
tags: 好观点 · → daily
🔴 bearish 侧:
- <a id="atom-0b0774ee5dd3bab2"></a>2026-06-13
X·@zephyr_z9 · 蒸馏已不再驱动能力提升 · 技术路线
Distillation has stopped being a capability driver for GLM, Deepseek, Moonshot
tags: 好观点 · → daily
- <a id="atom-65b3648a16c0e087"></a>2026-06-22
X·@_xjdr [老化中] · gap is larger for medical/legal domains · 技术路线_
further away for things like medical / legal / etc
tags: 好观点 · → daily
value: direction=wider gap
- <a id="atom-3aece742aea68af7"></a>2026-06-23
X·@teortaxesTex · lacks knowledge to train at 'Fable level' · 技术路线
but they don't know how
tags: 好观点 · → daily
展开 14 条中性
- <a id="atom-5aea8a0c9781eab7"></a>2026-05-21
X·@Sino_Market · 5.1 API发布 · 产品发布/技术路线
GLM-5.1 API LAUNCH
tags: 好信源 · → daily
📷 原图
- <a id="atom-332ac7762cbae30b"></a>2026-06-17
X·@RedRibbon131420 · 开源模型可能被削弱 · 技术路线
他们的开源模型被削弱的可能性
tags: 好问题 · → daily
- <a id="atom-6d3b5bdf84736853"></a>2026-06-18
X·@teortaxesTex · coder capability score range · 技术路线
Depending on task, lands between 4.6 and 4.8
tags: 好数字·好观点 · → daily
value: qty=4.6 to 4.8
- <a id="atom-5768127f16c946ea"></a>2026-06-19
X·@teortaxesTex · distillation from Claude indicated · 技术路线
to be clear this is happening partially because this is a claude harness
tags: 好观点 · → daily
- <a id="atom-d25200ecdd8684a2"></a>2026-06-18
X·@enolan · distilled from Claude · 技术路线
almost certainly distilled from Claude: thinks it's Claude by default and has extreme Claude voice
tags: 好观点 · → daily
2. 推理效率与缓存架构瓶颈
*topic: 推理效率/缓存命中率/基础设施缺陷 · 1 bull vs 5 bear*
🟢 bullish 侧:
- <a id="atom-03deab3b432fcabf"></a>2026-06-27
X·@scaling01 · 推理效率排名(高于DeepSeek) · 推理效率
_so reasoning efficiency:
Kimi > GLM > DeepSeek_
tags: 好观点 · → daily
🔴 bearish 侧:
- <a id="atom-b8d1fc81e4fd9c1b"></a>2026-05-09
X·@ZhihuFrontier · KV Cache hit rate after 2 minutes · 推理效率/缓存命中率 +同日4条
GLM 2min: 80% hit rate
tags: 好数字·好信源 · → daily
value: qty=80%
[图: 大语言模型API Token使用量与缓存命中情况的数据面板图表 — 总Token数: 50,949,110; 2026-04-28单日Token数: 21,213,830; 输入(命中缓存): 20,526,208; 输入(未命中缓存): 443,048; 输出: 244,574]
📷 原图
3. 监管限制削弱核心技术能力
*topic: 监管政策/模型发布 · 22 bull vs 3 bear*
🟢 bullish 侧:
- <a id="atom-c62b73f34d4f0b6f"></a>2026-05-25
X·@AiBattle_ · 新模型变体发布 · 模型发布
New “GLM-GA” model variant appears to be coming soon
tags: 好观点 · → daily
📷 原图
- <a id="atom-f7d7afcce101a5f7"></a>2026-06-13
X·@zephyr_z9 · 版本发布 · 模型发布
GLM 5.2 dropped at the right time
tags: 好观点 · → daily
value: date=2026-06-13
📷 原图
- <a id="atom-92affd21edb2f197"></a>2026-06-16
X·@teortaxesTex · 5.2 版本具备 COPE 能力 · 模型发布
GLM 5.2 also has… cheek. it even has COPE CAPACITY
tags: 好观点 · → daily
📷 原图
- <a id="atom-0bdd99b0d8bede77"></a>2026-06-17
X·@EMostaque · model quality · 模型发布
they have the other bits as GLM 5.2 shows
tags: 好观点 · → daily
- <a id="atom-4b4fb00827583e04"></a>2026-06-17
X·@teortaxesTex · 模型能力达到 latest-gen-Opus 水平 · 模型发布
我没有想到能达到最新一代 Opus 的水平
tags: 好观点 · → daily
🔴 bearish 侧:
- <a id="atom-b24a7a89f9812552"></a>2026-06-13
X·@AlphaExponent · 禁令情景提问 · 监管政策
Is there a scenario where they just ban GLM 5.2?
tags: 好问题 · → daily
value: direction=ban
- <a id="atom-cf3805e2635de2c9"></a>2026-06-17
X·@xhyctf · 团队表示已在训练期间移除网络安全能力 · 核心技术安全/模型发布
sadly, at a glm dev meetup just a few hours ago, the team told us straight up that they’d already taken out its cyber security during training.
tags: 好信源·好观点 · → daily
value: date=2026-06-17
- <a id="atom-83ba77c4a3c1ce5f"></a>2026-06-17
X·@teortaxesTex · 开源模型在网络相关方面被削弱 · 监管政策
他们在网络安全方面被削弱了
tags: 好信源 · → daily
展开 8 条中性
- <a id="atom-6e073cccdc757952"></a>2026-06-15
X·@AiBattle_ · API 发布时间 · 模型发布
GLM 5.2 API release is tomorrow, according to OpenRouter
tags: 好信源 · → daily
value: date=2026-06-16 · qty=tomorrow
📷 原图
- <a id="atom-62709ab1ee8377cb"></a>2026-06-18
X·@teortaxesTex · GLM 5.1 benchmark score · 模型发布
GLM 5.1 scores zero btw
tags: 好数字 · → daily
value: qty=0
- <a id="atom-4dd48c26c61d4314"></a>2026-06-21
X·@teortaxesTex · base model iterations pace · 模型发布/技术路线
GLM has an unusual pace in base model iterations, which is a bit obscured by versioning. GLM-4, 4.1, 4.5 are all different pretrains.
tags: 好观点 · → daily
value: qty=GLM-4, 4.1, 4.5 are all different pretrains
📷 原图
- <a id="atom-bcb6be9407e605df"></a>2026-06-25
X·@andonlabs · Vending-Bench 排名 · 模型发布
GLM 5.2 is 2nd in Vending-Bench.
tags: 好数字 · → daily
value: qty=2nd
[图: 不同AI模型在Vending-Bench 2模拟中资金余额随时间变化的对比图表 — 最高模型余额: 约$11,000; GLM-5.2最终余额: 约$8,200; GPT-5.5最终余额: 约$7,500; GLM-5.1最终余额: 约$5,600; GLM-5最终余额: 约$4,400]
📷 原图
- <a id="atom-e70f417dd74c4ab5"></a>2026-06-26
X·@teortaxesTex [老化中] · GLM 5.3 与 Opus 4.6/4.8 在复杂推理问题上的表现对比 · 技术路线/模型发布
Kind of questions where GLM 5.2 trips over (so does Opus 4.6, only Opus 4.8 does a pretty good job)
tags: 好观点 · → daily
4. 芯片约束与供给受限影响竞争力
*topic: 供给产能/竞争格局 · 14 bull vs 7 bear*
🟢 bullish 侧:
- <a id="atom-e7aad55672924993"></a>2026-05-09
X·@teortaxesTex [老化中] · 与顶级实验室的竞争力对比 · 模型性能/竞争格局
Kimi and GLM are actually competitive with top labs up to 128K
tags: 好观点 · → daily
- <a id="atom-3d59d86022b823d7"></a>2026-06-13
X·@xhyctf · 模型版本提升 · 模型发布/竞争格局
glm5.2 这次提升非常巨大远超 k2.7
tags: 好观点·好思考 · → daily
value: date=5.2
- <a id="atom-96b7a6be2344499b"></a>2026-06-17
X·@hxiao · GLM5.2 优于 MiniMax 同代模型 · 竞争格局
GLM5.2 blows MiniMax's same-gen model out of the water
tags: 好观点 · → daily
- <a id="atom-9a0d178cbcd05a4e"></a>2026-06-17
X·@teortaxesTex · 表现优于整个欧洲 · 竞争格局
他们比整个欧洲大陆都强
tags: 好观点 · → daily
📷 原图
- <a id="atom-b02cfb8abfa45d9a"></a>2026-06-18
X·@teortaxesTex · token efficiency vs DeepSeek vs Kimi · 技术路线/竞争格局 +同日1条
GLM is token-efficient compared to DS, Kimi
tags: 好观点 · → daily
📷 原图
🔴 bearish 侧:
- <a id="atom-9a10f900a31c7a26"></a>2026-05-22
X·@manateelazycat · 受 DeepSeek 降价影响 · 竞争格局
反而受到影响的是国内几家,比如Qwen、GLM、MiniMax 等
tags: 好观点 · → daily
value: statement=受到影响
📷 原图
- <a id="atom-919e74aaf62d3919"></a>2026-06-16
X·@Deepak_Reddy17 · 成本高于其他模型 · 成本/竞争格局
glm 5.2 I think will be costlier than any other model. Just saw one benchmark where it has to outspend 1:1 in terms of cost with opus for lower performance despite being more than 5x cheaper, with openai cost ratio is 7:1.
tags: 好观点·好问题 · → daily
- <a id="atom-4d423469aae939f4"></a>2026-06-17
X·@teortaxesTex · 在特定任务上需用2万token才近似正确 · 竞争格局/技术路线
GLM takes 20 thousand tokens to ALMOST get it exactly right, from first principles, fumbling in its CoT.
tags: 好观点 · → daily
value: qty=20000 tokens
📷 原图
- <a id="atom-9174fdab7e83cd13"></a>2026-06-21
X·@FredaDuan · 芯片约束 · 供给产能/技术路线 +同日1条
GLM is very chip constrained.
tags: 好思考 · → daily
- <a id="atom-4f938d50cbe57ca3"></a>2026-06-27
X·@teortaxesTex · rumors of being propped up by distillation or Claude Code interception · 技术路线/竞争格局
recent rumors about GLM being propped up by distillation (or rather, Claude Code use interception) are correct
tags: 好信源 · → daily
展开 5 条中性
- <a id="atom-c7fe0b64de1a98f2"></a>2026-06-13
X·@xhyctf · 产品推荐 · 产品发布/供给产能
如果打算用来做事的话推荐买 pro,lite 的额度还是太少了
tags: 好观点 · → daily
value: date=5.2
- <a id="atom-00803658d057e578"></a>2026-06-17
X·@teortaxesTex · 是中国最大的 LLM 创业公司 · 竞争格局
他们是中国最大的 LLM 创业公司
tags: 好信源 · → daily
📷 原图
- <a id="atom-4a687386592b0bcc"></a>2026-06-18
X·@teortaxesTex · reasoning tier comparison · 技术路线/竞争格局
I'd say Opus 4.7 tier but not as repulsive a personality
tags: 好观点 · → daily
🔗 因果传导 (causal map)
*GLM 在产业链上的传导关系. 边是群里/卖方陈述的因果 (非 AI 推断), 数字 = 几条 atom 支撑. 点 atom 溯源.*
↓ 下游·近期陈述 (1)
- ⚪ other Chinese labs · 份额外溢 · 1 次陈述
other Chinese labs can distill from GLM, now that it is a very effective model for SWE tasks
🟦 客观事实 (facts) (30)
- <a id="atom-4bf026970b2cbd5d"></a>🟦 2026-06-20
X·@teortaxesTex · GLM-5.2 API成本 · 价格动态
GLM-5.2 API调用成本: $59.60
tags: 好数字·好信源 · → daily
value: qty=$59.60
[图: Opus 4.8 与 GLM-5.2 的 API 调用成本对比柱状图 — Opus 4.8 成本: $183.72; GLM-5.2 成本: $59.60]
📷 原图
- <a id="atom-a45ac63395d7ee38"></a>🟦 2026-06-16
X·@Deepak_Reddy17 · 与opuse成本效率对比 · 成本
Just saw one benchmark where it has to outspend 1:1 in terms of cost with opus for lower performance despite being more than 5x cheaper.
tags: 好数字·好信源 · → daily
- <a id="atom-b8d1fc81e4fd9c1b"></a>🟦 2026-05-09
X·@ZhihuFrontier · KV Cache hit rate after 2 minutes · 推理效率/缓存命中率
GLM 2min: 80% hit rate
tags: 好数字·好信源 · → daily
value: qty=80%
[图: 大语言模型API Token使用量与缓存命中情况的数据面板图表 — 总Token数: 50,949,110; 2026-04-28单日Token数: 21,213,830; 输入(命中缓存): 20,526,208; 输入(未命中缓存): 443,048; 输出: 244,574]
📷 原图
- <a id="atom-8dbdefba9edc66d7"></a>🟦 2026-05-09
X·@ZhihuFrontier · KV Cache hit rate after 3 minutes · 推理效率/缓存命中率
GLM 3min: 50% hit rate
tags: 好数字·好信源 · → daily
value: qty=50%
[图: 大语言模型API Token使用量与缓存命中情况的数据面板图表 — 总Token数: 50,949,110; 2026-04-28单日Token数: 21,213,830; 输入(命中缓存): 20,526,208; 输入(未命中缓存): 443,048; 输出: 244,574]
📷 原图
- <a id="atom-0f075042c7b88fca"></a>🟦 2026-05-09
X·@ZhihuFrontier · KV Cache hit rate after 5 minutes · 推理效率/缓存命中率
GLM 5min: only 25% hit rate
tags: 好数字·好信源 · → daily
value: qty=25%
[图: 大语言模型API Token使用量与缓存命中情况的数据面板图表 — 总Token数: 50,949,110; 2026-04-28单日Token数: 21,213,830; 输入(命中缓存): 20,526,208; 输入(未命中缓存): 443,048; 输出: 244,574]
📷 原图
- <a id="atom-493b7c9851ae89b8"></a>🟦 2026-06-13
X·@Xianbao_QIAN · 5.2 发布 · 模型发布
GLM 5.2 is coming next week!
tags: 好信源·好思考 · → daily
value: date=2026-06-20
📷 原图
- <a id="atom-191c74331be1e73b"></a>🟦 2026-07-02
X·@thoughtfullab · inference cost comparison with Opus 4.8 · 价格动态/技术路线
GLM 5.2 is 5x cheaper than Opus 4.8
tags: 好数字·好观点 · → daily
value: qty=5x cheaper · date=2026-07-02
📷 原图
- <a id="atom-6b2b9d2a5ef69903"></a>🟦 2026-07-02
X·@thoughtfullab · inference cost comparison with Fable 5 · 价格动态/技术路线
GLM 5.2 is 11x cheaper than Fable 5
tags: 好数字·好观点 · → daily
value: qty=11x cheaper · date=2026-07-02
📷 原图
- <a id="atom-cf3805e2635de2c9"></a>🟦 2026-06-17
X·@xhyctf · 团队表示已在训练期间移除网络安全能力 · 核心技术安全/模型发布
sadly, at a glm dev meetup just a few hours ago, the team told us straight up that they’d already taken out its cyber security during training.
tags: 好信源·好观点 · → daily
value: date=2026-06-17
- <a id="atom-13675c2811621f1b"></a>🟦 2026-07-02
X·@teortaxesTex · 推理速度优化 · 模型发布/技术路线
This is the first DSpark speculator for a non-DeepSeek frontier model, trained with Speculators and running on vLLM nightly for ~1.5× faster d
tags: 好数字 · → daily
value: qty=1.5× · date=2026-07-02
📷 原图
- <a id="atom-39fae1b7778162b9"></a>🟦 2026-07-02
X·@pstAsiatech · 模型发布 · 模型发布
GLM 5.2 DSpark preview is here! ✨
tags: 好数字 · → daily
- <a id="atom-42aad0c63b04891e"></a>🟦 2026-06-27
X·@teortaxesTex · 速度 · 模型发布/技术路线
You can run GLM at 380
tags: 好数字 · → daily
value: qty=380 · unit=tps · direction=positive
- <a id="atom-fb839faff8c982df"></a>🟦 2026-06-27
X·@teortaxesTex · benchmark score · 模型发布/技术路线
Kimi K2.5 > GLM 5.2 and Opus 4.7
tags: 好数字 · → daily
value: qty=5.2
- <a id="atom-bf67088845aebdf8"></a>🟦 2026-06-26
X·@brightmirror · 第6代模型发布 · 模型发布_
Using GLM 6
tags: 好数字 · → daily
- <a id="atom-bcb6be9407e605df"></a>🟦 2026-06-25
X·@andonlabs · Vending-Bench 排名 · 模型发布
GLM 5.2 is 2nd in Vending-Bench.
tags: 好数字 · → daily
value: qty=2nd
[图: 不同AI模型在Vending-Bench 2模拟中资金余额随时间变化的对比图表 — 最高模型余额: 约$11,000; GLM-5.2最终余额: 约$8,200; GPT-5.5最终余额: 约$7,500; GLM-5.1最终余额: 约$5,600; GLM-5最终余额: 约$4,400]
📷 原图
- <a id="atom-f5128069af760b8c"></a>🟦 2026-06-25
X·@andonlabs · 每版本改进幅度 · 技术路线
Each GLM release has improved at a remarkably steady pace: a linear fit of R²=0.99, with almost $1k better per month.
tags: 好数字 · → daily
value: qty=$1k per month
[图: 不同AI模型在Vending-Bench 2模拟中资金余额随时间变化的对比图表 — 最高模型余额: 约$11,000; GLM-5.2最终余额: 约$8,200; GPT-5.5最终余额: 约$7,500; GLM-5.1最终余额: 约$5,600; GLM-5最终余额: 约$4,400]
📷 原图
- <a id="atom-8a0ef22f7b09400b"></a>🟦 2026-06-24
X·@teortaxesTex · ARC-AGI-2得分 · 模型发布/技术路线
GLM 5.2 is the best Chinese model on ARC-AGI-2, at 22.8%
tags: 好数字 · → daily
value: qty=22.8% · direction=
📷 原图
- <a id="atom-3d05b19f4bc9c191"></a>🟦 2026-06-21
X·@teortaxesTex · scale of training models · 技术路线/供给产能
tinker with 32B llama-likes, then moving onto a 355B MoE, then to 744B DeepSeek clone
tags: 好数字 · → daily
value: qty=from 32B to 355B MoE to 744B DeepSeek clone
- <a id="atom-dab30d51f9c33cf8"></a>🟦 2026-06-20
X·@mweinbach · API价格 · API价格
$200 of GLM 5.2 API usage
tags: 好数字 · → daily
value: qty=$200 · unit=$200 of GLM 5.2 API usage
- <a id="atom-dc56db95b57bea32"></a>🟦 2026-06-20
X·@mweinbach · 订阅价格历史 · 订阅价格
GLM Coding Pro 续费金额: $90
tags: 好数字 · → daily
value: qty=$90 · unit=GLM Coding Pro 续费金额: $90
[图: GLM Coding 订阅服务的历史订单交易记录数据面板 — GLM Coding Pro 续费金额: $90; GLM Coding Pro 升级金额: $45; GLM Coding Lite 支付金额: $3]
📷 原图
- <a id="atom-147d562bb24352cb"></a>🟦 2026-06-20
X·@mweinbach · 订阅升级价格 · 订阅价格
GLM Coding Pro 升级金额: $45
tags: 好数字 · → daily
value: qty=$45 · unit=GLM Coding Pro 升级金额: $45
[图: GLM Coding 订阅服务的历史订单交易记录数据面板 — GLM Coding Pro 续费金额: $90; GLM Coding Pro 升级金额: $45; GLM Coding Lite 支付金额: $3]
📷 原图
- <a id="atom-1787a4d95fef6a3e"></a>🟦 2026-06-20
X·@JordanNanos · API 输入价格 · 价格动态
GLM 5.2 costs $1.40/4.40 per Mtok at 40 tok/sec
tags: 好数字 · → daily
value: qty=$1.40/Mtok · currency=USD · unit=per Mtok
[图: 不同大模型API服务商的Token输入与输出价格对比数据面板 — Wafer输入价格: $1.20/M; Wafer输出价格: $4.10/M; 多数服务商输入价格: $1.40/M; 多数服务商输出价格: $4.40/M; io.net输出价格: $5.28/M]
📷 原图
- <a id="atom-1ae27d8aab696cf9"></a>🟦 2026-06-20
X·@JordanNanos · API 输出价格 · 价格动态
GLM 5.2 costs $1.40/4.40 per Mtok at 40 tok/sec
tags: 好数字 · → daily
value: qty=$4.40/Mtok · currency=USD · unit=per Mtok
[图: 不同大模型API服务商的Token输入与输出价格对比数据面板 — Wafer输入价格: $1.20/M; Wafer输出价格: $4.10/M; 多数服务商输入价格: $1.40/M; 多数服务商输出价格: $4.40/M; io.net输出价格: $5.28/M]
📷 原图
- <a id="atom-ed26e27ad664cde5"></a>🟦 2026-06-20
X·@TheAhmadOsman · 模型参数量 · 模型发布
Luke Alonso has uploaded an NVFP4 of GLM 5.2
tags: 好数字 · → daily
value: qty=5.2 · date=2026-06-20 · direction=na
📷 原图
- <a id="atom-62709ab1ee8377cb"></a>🟦 2026-06-18
X·@teortaxesTex · GLM 5.1 benchmark score · 模型发布
GLM 5.1 scores zero btw
tags: 好数字 · → daily
value: qty=0
- <a id="atom-a34bc00f217ac663"></a>🟦 2026-06-18
X·@Zai_org · 5.2 app development benchmark score · 模型发布
GLM-5.2: 48/70
tags: 好数字 · → daily
value: qty=48/70
- <a id="atom-78aa41043f01824b"></a>🟦 2026-06-17
X·@teortaxesTex · 开发机构 · 合作客户
GLM has been developed by Tsinghua University Data Mining Group
tags: 好信源 · → daily
📷 原图
- <a id="atom-267d0b7ba6d8b5c7"></a>🟦 2026-06-17
X·@teortaxesTex · 推理成本对比 · 模型发布/技术路线
GLM 4.7 is about 2x smaller than GLM-5.2
tags: 好数字 · → daily
value: qty=2x · direction=smaller
- <a id="atom-b6304e0e1725aad5"></a>🟦 2026-06-16
X·@Deepak_Reddy17 · 与openai成本效率对比 · 成本
with openai cost ratio is 7:1.
tags: 好数字 · → daily
- <a id="atom-a192f6f811ca1fc6"></a>🟦 2026-06-15
X·@ZhihuFrontier · GLM-4.7 参数量 · 参数量
GLM-4.7: 358B parameters
tags: 好数字 · → daily
value: qty=358B
📷 原图
🟥 多头 takes (bullish) (30)
- <a id="atom-b2812fd8a72a9295"></a>🟥 2026-06-22
X·@AravSrinivas · 参数规模 · 模型发布/技术路线
Is a sub trillion parameter model, meaning it has a lot of potential to go beyond matching the frontier at the median level of difficulty to also doing it for the long tail.
tags: 好观点·好思考 · → daily
value: direction=sub trillion parameter
- <a id="atom-b8a34c6170ae2182"></a>🟥 2026-06-21
X·@rationaleist · general capabilities jump from 5.1 to 5.2 · 模型发布/技术路线
I'm just really impressed by 5.1 -> 5.2 jump on general capabilities ig. I expected code but not everything else. Feels like they cracked something.
tags: 好观点·好思考 · → daily
- <a id="atom-3d59d86022b823d7"></a>🟥 2026-06-13
X·@xhyctf · 模型版本提升 · 模型发布/竞争格局
glm5.2 这次提升非常巨大远超 k2.7
tags: 好观点·好思考 · → daily
value: date=5.2
- <a id="atom-21fcc4cb02ffb152"></a>🟥 2026-06-23
X·@teortaxesTex · 推理token效率对比 · 技术路线
GLM is far better, 1/5 tokens (so same cost and ≈3x faster)
tags: 好数字 · → daily
value: qty=1/5 tokens · date=2026-06-23
📷 原图
- <a id="atom-287cf0554ec880ed"></a>🟥 2026-06-19
X·@teortaxesTex · 模型发布间隔 · 模型发布/技术进步
GLM-5.1 came out 2 weeks before V4, and was already stronger. they don't have GLM-5 scores in any case, on its own this indicates a great acceleration in Zhipu program.
tags: 好数字 · → daily
value: qty=2周 · direction=加速
📷 原图
- <a id="atom-d5df06ce5b9738fb"></a>🟥 2026-06-27
X·@sir_deenicus · performance on challenging codebases · 产品发布/竞争格局
It's only open model competent on one of my more challenging codebases
tags: 好信源 · → daily
📷 原图
- <a id="atom-03deab3b432fcabf"></a>🟥 2026-06-27
X·@scaling01 · 推理效率排名(高于DeepSeek) · 推理效率
_so reasoning efficiency:
Kimi > GLM > DeepSeek_
tags: 好观点 · → daily
- <a id="atom-65f184b35e704e4b"></a>🟥 2026-06-13
X·@AiBattle_ [老化中] · 5.2 is a significant improvement over 5.1 · 产品发布/技术路线
GLM 5.2 seems to be a significant improvement over GLM 5.1 and has a Max Thinking option
tags: 好观点·好问题 · → daily
📷 原图
- <a id="atom-ef784b9ec0eafa00"></a>🟥 2026-06-29
X·@ii_posts · 开源模型性能 · 技术路线/模型发布
GLM 5.2 通过自适应自我改进将基础模型提升到 FrontierSWE 顶端
tags: 好思考 · → daily
value: qty=FrontierSWE top 1
📷 原图
- <a id="atom-18107c9739dd16de"></a>🟥 2026-06-26
X·@myainotez · 模型能力与GPT 5.2对比 · 模型发布/技术路线
glm 5.2 will do numbers. Even on kernel topics you can make it on par with gpt 5.5 high
tags: 好思考 · → daily
- <a id="atom-fe4b9cc4309e5db6"></a>🟥 2026-06-25
X·@Xianbao_QIAN · 性能对比 · 模型发布/竞争格局
comparable with GLM5.2
tags: 好思考 · → daily
value: qty=comparable · direction=up
📷 原图
- <a id="atom-2369ff6b15d8e051"></a>🟥 2026-06-24
X·@TheAhmadOsman · 性能超过 GPT 5.5 · 技术路线
GLM 5.5 > GPT 5.5 XHIGH
tags: 好思考 · → daily
value: qty=XHIGH
- <a id="atom-b8a1d6c9c8a5cc4e"></a>🟥 2026-06-22
X·@interconnectsai · 5.2 is the step change for open agents · 模型发布/技术路线
GLM-5.2 is the step change for open agents
tags: 好思考 · → daily
- <a id="atom-c83df4e35adf238b"></a>🟥 2026-06-17
X·@teortaxesTex · 模型计划添加视觉能力 · 产品发布
vision is very important, I constantly see DeepSeek stumbling trying to use vision tools in harnesses
tags: 好思考 · → daily
- <a id="atom-a262703e39efcb28"></a>🟥 2026-06-15
X·@ZhihuFrontier · Chain-of-Thought 结构特点 · 推理能力
GLM's Chain-of-Thought has a surprisingly consistent structure. Extract every key requirement, summarize each point explicitly, recursively compress and reorganize information
tags: 好思考 · → daily
📷 原图
- <a id="atom-c8ff03008c780e21"></a>🟥 2026-06-15
X·@ZhihuFrontier · 通过推理压缩信息假设 · 推理能力
GLM seems to use reasoning to compress information. Each round produces a smaller, more structured representation of the problem
tags: 好思考 · → daily
📷 原图
- <a id="atom-09aff98bd13705fa"></a>🟥 2026-07-01
X·@teortaxesTex · coding eval coverage · 技术路线
GLM covered most of that gap
tags: 好观点 · → daily
📷 原图
- <a id="atom-b198851b32018a92"></a>🟥 2026-06-29
X·@dhtikna · has better chance if post training is key · 技术路线
if post training is the key then I think glm or kimi have a better chance
tags: 好观点 · → daily
- <a id="atom-9202cc912807ca02"></a>🟥 2026-06-27
X·@TheAhmadOsman [老化中] · 订阅推荐 · 产品发布
GLM / Kimi subscriptions are decent as well
tags: 好观点 · → daily
- <a id="atom-4309f37a811ed389"></a>🟥 2026-06-26
X·@bookwormengr [老化中] · 700B 参数规模下的表现被认为 exceptional · 技术路线/供给产能
On the other hand what GLM has done with 700B parameters is exceptional.
tags: 好观点 · → daily
- <a id="atom-aecf8eb4bfaf9299"></a>🟥 2026-06-25
X·@teortaxesTex · 领先其他中国模型 · 竞争格局
GLM is far ahead of other Chinese models now, and certainly Deepseek
tags: 好观点 · → daily
- <a id="atom-dee21f39d6d0c7c3"></a>🟥 2026-06-25
X·@bookwormengr · 5.2 版本在敏感问题回答上表现良好 · 技术路线/模型发布
Asking below the sensitive question around Tianmen Square events in 1989. This is on an API hosted in the USA. This is the hugging-face version, I suppose. Not sure what the complaint against them is at least on this ground.
tags: 好观点 · → daily
value: date=2026-06-25
📷 原图
- <a id="atom-de89350cb4620c06"></a>🟥 2026-06-25
X·@max_paperclips · PPO research quality · 模型发布/技术路线
GLM doing GOOD long horizon PPO
tags: 好观点 · → daily
value: direction=good long horizon
- <a id="atom-a1b86cd0d47aa424"></a>🟥 2026-06-23
X·@teortaxesTex · compute sufficient to train at 'Fable level' · 供给产能
Multiple Chinese entities have the compute to train 'Fable level'
tags: 好观点 · → daily
- <a id="atom-44a33b7f469e322a"></a>🟥 2026-06-22
X·@teortaxesTex [老化中] · holds up on legal benchmarks · 技术路线
GLM kinda holds up even on legal, it's surprising.
tags: 好观点 · → daily
- <a id="atom-7d1b464f68f5e2b3"></a>🟥 2026-06-22
X·@AravSrinivas · opensource AI 模型 · 技术路线/模型发布
GLM is the kind of model that revives serious interest in open source AI.
tags: 好观点 · → daily
- <a id="atom-4ce606fb82067c8f"></a>🟥 2026-06-22
X·@AravSrinivas · 表现 · 模型发布/竞争格局
It passes the blind test relative to the frontier models on the median production grade knowledge worker task.
tags: 好观点 · → daily
value: direction=passes blind test relative to frontier models on median production grade knowledge worker task
- <a id="atom-0a92d49463b89b78"></a>🟥 2026-06-22
X·@AravSrinivas · 推理成本 · 定价权/竞争格局
It’s affordable to serve.
tags: 好观点 · → daily
value: direction=affordable to serve
- <a id="atom-1e266ff1bab755d9"></a>🟥 2026-06-21
X·@teortaxesTex · 5th gen model timing · 模型发布
I can see them milking 5th gen to Q4 2026 and making a jump.
tags: 好观点 · → daily
value: date=Q4 2026
📷 原图
- <a id="atom-a006dc33c61dfa98"></a>🟥 2026-06-21
X·@FredaDuan · 云与本地计算效率对比 · 价格动态/数据中心算力
Cloud compute remains more token-efficient per total cost dollar than local compute.
tags: 好观点 · → daily
🟥 空头 takes (bearish) (17)
- <a id="atom-7dcc41316d743d3d"></a>🟥 2026-06-28
X·@negligible_cap · 模型能力对比 · 模型发布/竞争格局
GLM-5.2 may have created another DeepSeek moment according to Jeffries: this new model is almost equal to Anthropic as a competitor for the corporate market and is just one quarter of the cost in terms of cost per token
tags: 好信源·好思考 · → daily
📷 原图
- <a id="atom-e644727e93fafc42"></a>🟥 2026-06-21
X·@FredaDuan · 本地部署成本效益 · 资本开支/定价权
If it takes $20k of hardware to run GLM 5.2 and you only break even after 5.5 years of 24/7 utilization, cloud still wins.
tags: 好数字·好思考 · → daily
value: qty=$20k · date=5.5 years
- <a id="atom-a33de15bdb7e5ac1"></a>🟥 2026-05-09
X·@ZhihuFrontier · infrastructure architecture defect · 基础设施缺陷/缓存架构
Infra architecture defect: Unable to offload KV Cache to low-cost disk storage, strictly limited to on-board VRAM. Small cache pool forces aggressive LRU eviction.
tags: 好思考·好信源 · → daily
[图: 大语言模型API Token使用量与缓存命中情况的数据面板图表 — 总Token数: 50,949,110; 2026-04-28单日Token数: 21,213,830; 输入(命中缓存): 20,526,208; 输入(未命中缓存): 443,048; 输出: 244,574]
📷 原图
- <a id="atom-cf2bf3931a14eeec"></a>🟥 2026-05-09
X·@ZhihuFrontier [老化中] · extreme traffic throughput impact on KV Cache · 基础设施缺陷/缓存架构
Extreme traffic throughput far exceeds cache bearing capacity, accelerating invalidation of historical KV sequences.
tags: 好思考·好信源 · → daily
[图: 大语言模型API Token使用量与缓存命中情况的数据面板图表 — 总Token数: 50,949,110; 2026-04-28单日Token数: 21,213,830; 输入(命中缓存): 20,526,208; 输入(未命中缓存): 443,048; 输出: 244,574]
📷 原图
- <a id="atom-4f938d50cbe57ca3"></a>🟥 2026-06-27
X·@teortaxesTex · rumors of being propped up by distillation or Claude Code interception · 技术路线/竞争格局
recent rumors about GLM being propped up by distillation (or rather, Claude Code use interception) are correct
tags: 好信源 · → daily
- <a id="atom-919e74aaf62d3919"></a>🟥 2026-06-16
X·@Deepak_Reddy17 · 成本高于其他模型 · 成本/竞争格局
glm 5.2 I think will be costlier than any other model. Just saw one benchmark where it has to outspend 1:1 in terms of cost with opus for lower performance despite being more than 5x cheaper, with openai cost ratio is 7:1.
tags: 好观点·好问题 · → daily
- <a id="atom-0b0774ee5dd3bab2"></a>🟥 2026-06-13
X·@zephyr_z9 · 蒸馏已不再驱动能力提升 · 技术路线
Distillation has stopped being a capability driver for GLM, Deepseek, Moonshot
tags: 好观点 · → daily
- <a id="atom-0c02d2a93c017c66"></a>🟥 2026-06-25
X·@marcospereeira · believes it’s claude · 产品发布
GLM believes it’s claude
tags: 好观点 · → daily
- <a id="atom-3aece742aea68af7"></a>🟥 2026-06-23
X·@teortaxesTex · lacks knowledge to train at 'Fable level' · 技术路线
but they don't know how
tags: 好观点 · → daily
- <a id="atom-7f34590ced903d43"></a>🟥 2026-06-23
X·@thalevinod · 境外用户注册流程 · 用户体验/注册流程
It is extremely frustrating that foreigners in China cannot easily access GLM. Other coding agents, such as Kimi and Qwen, make registration, payment, and usage straightforward, but GLM has made the process much more difficult than it needs to be.
tags: 好观点 · → daily
value: direction=more difficult
- <a id="atom-65b3648a16c0e087"></a>🟥 2026-06-22
X·@_xjdr [老化中] · gap is larger for medical/legal domains · 技术路线_
further away for things like medical / legal / etc
tags: 好观点 · → daily
value: direction=wider gap
- <a id="atom-b45aa9528e43a586"></a>🟥 2026-06-20
X·@mweinbach · 代码计划价值 · 订阅价值
Their coding plans tend to be a worse value than others and also unbearably slow
tags: 好观点 · → daily
- <a id="atom-d5ccb88cc40f5c10"></a>🟥 2026-06-20
X·@mweinbach · 订阅价格变化趋势 · 订阅价格
it is no longer cheap
tags: 好观点 · → daily
[图: GLM Coding 订阅服务的历史订单交易记录数据面板 — GLM Coding Pro 续费金额: $90; GLM Coding Pro 升级金额: $45; GLM Coding Lite 支付金额: $3]
📷 原图
- <a id="atom-4d423469aae939f4"></a>🟥 2026-06-17
X·@teortaxesTex · 在特定任务上需用2万token才近似正确 · 竞争格局/技术路线
GLM takes 20 thousand tokens to ALMOST get it exactly right, from first principles, fumbling in its CoT.
tags: 好观点 · → daily
value: qty=20000 tokens
📷 原图
- <a id="atom-76e490911e188194"></a>🟥 2026-06-14
X·@_xjdr · 模型使用体验 · 模型性能_
similar to qwen, i've just never had real success with their models for real work
tags: 好观点 · → daily
- <a id="atom-9a10f900a31c7a26"></a>🟥 2026-05-22
X·@manateelazycat · 受 DeepSeek 降价影响 · 竞争格局
反而受到影响的是国内几家,比如Qwen、GLM、MiniMax 等
tags: 好观点 · → daily
value: statement=受到影响
📷 原图
- <a id="atom-b24a7a89f9812552"></a>🟥 2026-06-13
X·@AlphaExponent · 禁令情景提问 · 监管政策
Is there a scenario where they just ban GLM 5.2?
tags: 好问题 · → daily
value: direction=ban
🟥 中性 takes (neutral) (20)
- <a id="atom-8f5e61d78e08f2fd"></a>🟥 2026-06-21
X·@teortaxesTex · model scale requirement · 技术路线
They need to scale up at least to 2.4T I think
tags: 好观点·好数字 · → daily
value: qty=at least to 2.4T parameters
- <a id="atom-6d3b5bdf84736853"></a>🟥 2026-06-18
X·@teortaxesTex · coder capability score range · 技术路线
Depending on task, lands between 4.6 and 4.8
tags: 好数字·好观点 · → daily
value: qty=4.6 to 4.8
- <a id="atom-766b1e4ab560803b"></a>🟥 2026-06-26
X·@max_paperclips · 技术路线转变 · 技术路线
GLM have seen the light and returned to PPO
tags: 好观点 · → daily
- <a id="atom-daa908a355f999c2"></a>🟥 2026-06-21
X·@teortaxesTex · 另有预训练计划 · 技术路线
obviously not about 5.2. And that's interesting, it means they'll have another pretrain
tags: 好思考 · → daily
- <a id="atom-1975eac7633342d7"></a>🟥 2026-06-20
X·@kalomaze [老化中] · generalist tool use ability · 模型能力/技术路线
it's distinctly less generalist in terms of 'robust domain specific capabilities' but distinctly more generalist in terms of 'tool use ICL g factor executing long horizon agentic loops without exploding'
tags: 好思考 · → daily
- <a id="atom-368995949d043cea"></a>🟥 2026-07-01
X·@hxiao · vision integration urgency · 产品发布/技术路线
I don't see that super urgency to bake VLM into GLM next release
tags: 好观点 · → daily
- <a id="atom-e70f417dd74c4ab5"></a>🟥 2026-06-26
X·@teortaxesTex [老化中] · GLM 5.3 与 Opus 4.6/4.8 在复杂推理问题上的表现对比 · 技术路线/模型发布
Kind of questions where GLM 5.2 trips over (so does Opus 4.6, only Opus 4.8 does a pretty good job)
tags: 好观点 · → daily
- <a id="atom-12c8049a907000dd"></a>🟥 2026-06-22
X·@teortaxesTex [老化中] · gap to frontier equals approximately 7 months · 技术路线
I think GLM 5.2 makes the gap at present equal to roughly 7 months, all things considered.
tags: 好观点 · → daily
value: qty=7 months · direction=gap
- <a id="atom-b79e9f9045de93c6"></a>🟥 2026-06-22
X·@_xjdr · gap is roughly 7 months for coding · 技术路线_
~7 months seems reasonable tho (i'd say its closer for coding and further away for things like medical / legal / etc)
tags: 好观点 · → daily
value: qty=7 months · direction=gap
- <a id="atom-4dd48c26c61d4314"></a>🟥 2026-06-21
X·@teortaxesTex · base model iterations pace · 模型发布/技术路线
GLM has an unusual pace in base model iterations, which is a bit obscured by versioning. GLM-4, 4.1, 4.5 are all different pretrains.
tags: 好观点 · → daily
value: qty=GLM-4, 4.1, 4.5 are all different pretrains
📷 原图
- <a id="atom-5768127f16c946ea"></a>🟥 2026-06-19
X·@teortaxesTex · distillation from Claude indicated · 技术路线
to be clear this is happening partially because this is a claude harness
tags: 好观点 · → daily
- <a id="atom-5a5a826f241d50ad"></a>🟥 2026-06-19
X·@TheAhmadOsman · 智能水平 · 模型性能
GLM can be generally more intelligent when it doesn't fall into an overthinking loop
tags: 好观点 · → daily
- <a id="atom-4a687386592b0bcc"></a>🟥 2026-06-18
X·@teortaxesTex · reasoning tier comparison · 技术路线/竞争格局
I'd say Opus 4.7 tier but not as repulsive a personality
tags: 好观点 · → daily
- <a id="atom-2c41cb26ce906515"></a>🟥 2026-06-18
X·@teortaxesTex · previous tier classification · 技术路线/竞争格局
I had them pegged as tier 2 Chinese lab
tags: 好观点 · → daily
value: qty=tier 2
- <a id="atom-79e67c49c389076d"></a>🟥 2026-06-18
X·@teortaxesTex · tokenizer shared with Mythos · 技术路线
Because Mythos existed by then, and they share tokenizers
tags: 好观点 · → daily
- <a id="atom-d25200ecdd8684a2"></a>🟥 2026-06-18
X·@enolan · distilled from Claude · 技术路线
almost certainly distilled from Claude: thinks it's Claude by default and has extreme Claude voice
tags: 好观点 · → daily
- <a id="atom-27c2f7b58e50bdfe"></a>🟥 2026-06-13
X·@mranti [老化中] · 个人喜好 · 仓位情绪
我对GLM没那么爱。
tags: 好观点 · → daily
- <a id="atom-c7fe0b64de1a98f2"></a>🟥 2026-06-13
X·@xhyctf · 产品推荐 · 产品发布/供给产能
如果打算用来做事的话推荐买 pro,lite 的额度还是太少了
tags: 好观点 · → daily
value: date=5.2
- <a id="atom-789b99ecf4f28ec8"></a>🟥 2026-06-23
X·@pastaraspberry · 股价表现 · 价格动态
glm is flying tonight.
tags: 好问题 · → daily
value: qty=上涨 · date=2026-06-23
- <a id="atom-332ac7762cbae30b"></a>🟥 2026-06-17
X·@RedRibbon131420 · 开源模型可能被削弱 · 技术路线
他们的开源模型被削弱的可能性
tags: 好问题 · → daily
⏱ 时间轴 (近 20)
- 🟦 2026-07-02 ·
fact · 推理速度优化 · → daily
- 🟦 2026-07-02 ·
fact · 模型发布 · → daily
- 🟦 2026-07-02 ·
fact · inference cost comparison with Opus 4.8 · → daily
- 🟦 2026-07-02 ·
fact · inference cost comparison with Fable 5 · → daily
- 🟦 2026-07-01 ·
fact · 接近前沿模型水平,使用极少量受限制计算资源 · → daily
- 🟥 2026-07-01 ·
forecast · vision integration urgency · → daily
- 🟥 2026-07-01 ·
narrative · coding eval coverage · → daily
- 🟦 2026-07-01 ·
forecast · 模型发布计划 · → daily
- 🟦 2026-07-01 ·
forecast · 模型版本参数规模 · → daily
- 🟥 2026-06-29 ·
forecast · has better chance if post training is key · → daily
- 🟥 2026-06-29 ·
narrative · 开源模型性能 · → daily
- 🟥 2026-06-28 ·
narrative · 模型能力对比 · → daily
- 🟦 2026-06-27 ·
fact · 速度 · → daily
- 🟥 2026-06-27 ·
narrative · rumors of being propped up by distillation or Claude Code in · → daily
- 🟥 2026-06-27 ·
narrative · performance on challenging codebases · → daily
- 🟦 2026-06-27 ·
fact · benchmark score · → daily
- 🟦 2026-06-27 ·
fact · ablation · → daily
- 🟦 2026-06-27 ·
fact · weight quantization · → daily
- 🟥 2026-06-27 ·
narrative · 推理效率排名(高于DeepSeek) · → daily
- 🟥 2026-06-27 ·
position · 订阅推荐 · → daily
← 实体目录 · 系统日志