tphuang retweeted
This...
Kimi 3.1 is the best released Chinese model but progression over 3.0 in 3 months is underwhelming.
It’s Opus5.0/GPT-6 Sol level rather than Opus5.5/6.1 level. Not bad, but DeepSeek has clearly progressed faster with its RL than Moonshot. In cyber & some SWE tasks, 1017 is actuDS flash update will catch up to this on SWE & work tasks.
So then what do you do if you are Moonshot? Well, I wonder if it’s just pouring most of its energy on the larger K4. Among top Chinese labs, it seems best at pre training & worst @ post training.
So it’s best bet is keep scaling up. Its models are not efficient like DeepSeek models. It has never pretended otherwise. Even at current px level, it’s overwhelmed by demand so why should it charge less? You should be using DeepSeek if you are just doing regular work stuff.
0
1
8
1,705
tphuang retweeted
Kimi K3.1 Pro 猫老板评测文章出炉
先说结论：
后训练大参数模型踩坑了，有进步但不如预期
高情商「把大参数的坑都踩一遍」
低情商「做得和一坨一样，实在没办法夸了，就拿来当反面例子吧」
以下是快速总结：
K3.1 整体明显优于 K3，多个能力项达到国产 术栈和“徒手”建模是明显短板。
前端 UI🟢🟢🟢 国产第一，达到生产级
信息提取（纯）🟢🟢 接近 GPT，稳定性略差
复杂求解🟢🟢 接近 GPT，token 效率高
效率（熟悉领域）🟢🟢 与 GPT-6 持平
指令遵循🟡 小细节会漏掉，但换个说法就能沟通
不使用工具时的建模🔴 国内最强，但远不及 Opus
冷门技术栈🔴 需要反复查文档，会反复做低价值调整
适合的与不适合的场景：
✅ 适合：
前端/UI 开发（直出生产级）
复杂探索性任务（token 效率高）
审美要求高的界面
信息提取（纯文本类）
❌ 不适合
不使用工具的纯建模（除非你只要“看起来像”就行）
冷门技术栈的快速原型（会反复查文档）
需要零遗漏的细节指令（要反复强调）
带逻辑的数据提取（不如 DeepSeek Flash）
网友核心吐槽：
1.太贵 —— 「Kimi 的问题就是模型成本下不来，订阅用不了多少额度，对个人生产力性价比不高」
2.没达到预期 —— 「老外也说没达到预期」
3.后端平庸 —— 「除了前端之外，后端跟 GLM-5.3 一个水平，连 DeepSeek Flash 都不如」
4.算力不够还硬上 —— 「是不是就是单纯没法子了，只能堆参数」
先说结论：
后训练大参数模型踩坑了，有进步但不如预期
高情商「把大参数的坑都踩一遍」
低情商「做得和一坨一样，实在没办法夸了，就拿来当反面例子吧」
以下是快速总结：
K3.1 整体明显优于 K3，多个能力项达到国产 术栈和“徒手”建模是明显短板。
前端 UI🟢🟢🟢 国产第一，达到生产级
信息提取（纯）🟢🟢 接近 GPT，稳定性略差
复杂求解🟢🟢 接近 GPT，token 效率高
效率（熟悉领域）🟢🟢 与 GPT-6 持平
指令遵循🟡 小细节会漏掉，但换个说法就能沟通
不使用工具时的建模🔴 国内最强，但远不及 Opus
冷门技术栈🔴 需要反复查文档，会反复做低价值调整
适合的与不适合的场景：
✅ 适合：
前端/UI 开发（直出生产级）
复杂探索性任务（token 效率高）
审美要求高的界面
信息提取（纯文本类）
❌ 不适合
不使用工具的纯建模（除非你只要“看起来像”就行）
冷门技术栈的快速原型（会反复查文档）
需要零遗漏的细节指令（要反复强调）
带逻辑的数据提取（不如 DeepSeek Flash）
网友核心吐槽：
1.太贵 —— 「Kimi 的问题就是模型成本下不来，订阅用不了多少额度，对个人生产力性价比不高」
2.没达到预期 —— 「老外也说没达到预期」
3.后端平庸 —— 「除了前端之外，后端跟 GLM-5.3 一个水平，连 DeepSeek Flash 都不如」
4.算力不够还硬上 —— 「是不是就是单纯没法子了，只能堆参数」
10
2
47
11,881
Updated my interactive map. Now, you can see the 4 countries that Wang Yi visited + the North corridor & the original Middle Corridor.
"Plan B" passes through Caucasus -> Black Sea ->Bulgaria -> Serbia -> Croatia.
Goal is decrease leverage/risk wrt Russia, Poland & Turkey.
Brings increased emnvolved like Kazakhstan, Azerbaijan, Bulgaria & Croatia.
"Plan B" passes through Caucasus -> Black Sea ->Bulgaria -> Serbia -> Croatia.
Goal is decrease leverage/risk wrt Russia, Poland & Turkey.
Brings increased emnvolved like Kazakhstan, Azerbaijan, Bulgaria & Croatia.
DeepSeek V4.1-1017 is too bad-ass.
I've wanted an interactive map forever.
Now, I just put a prompt into DSH & 43min later, I got my interactive map to show China's new Middle Corridor alternative that bypasses Turkey & goes thru Caucasus & Bulgaria to reach Budapest.
Cost $0.48
I know, youthis w/ your beloved Opus or GPT6.1-Sol, but would it have only cost 48 cent?
That's the power of DeepSeek now. You know it will get better every month. You know DSH keeps getting better. And you know it's basically free, so even El Cheapos like me can afford it.
Basically, the limitation now is fully on me, not on DeepSeek.
3
9
66
3,374
China's import & Export of Tungsten in 2026 vs 2025:
Huge increase in W concentrate/ore as it continues to increase refining capacity.
Decreased export in volume of Unwrought W, W powder & Ferrotungsten.
Exporting less intermediate W & more downstream + higher valued W product
Huge increase in W concentrate/ore as it continues to increase refining capacity.
Decreased export in volume of Unwrought W, W powder & Ferrotungsten.
Exporting less intermediate W & more downstream + higher valued W product
1
6
43
2,485
BYD's plant in Bahia Brazil plans to raise production capacity to 600k in 2027 & reach 50% local rate by early 2027 to be considered manufactured locally.
From there, it can be exported to rest of Latam, including Argentina & Mexico, which would otherwise have tariffs on Chinese cars.
From there, it can be exported to rest of Latam, including Argentina & Mexico, which would otherwise have tariffs on Chinese cars.
2
11
84
5,038
BYD Semiconductor announces reaching 50m delivery of its isolated gate driver chips BF1190 & BF1186 series.
Sic die -> module packaging -> BMC AFE -> Motor MCU. Isolated drivers is also bw MCU & power devices.
Used in BYD effort to go from 400V to 1000V platform for its EVs.
Sic die -> module packaging -> BMC AFE -> Motor MCU. Isolated drivers is also bw MCU & power devices.
Used in BYD effort to go from 400V to 1000V platform for its EVs.
1
7
90
3,300
Looks like tanker strikes are finally really making things uncomfortable for Emiratis to the pt they are now paying Iranians for transits. Physical oil is going way up. Gas px in great plains up bigly due to refinery issues.
It looks like Trump is getting ready to go in. It's bw he US empire. I don't see how he can choose the latter this early on. No empire has given up on its global hegemony bc prices got high
It looks like Trump is getting ready to go in. It's bw he US empire. I don't see how he can choose the latter this early on. No empire has given up on its global hegemony bc prices got high
1
9
111
5,543
DeepSeek V4.1-1017 is too bad-ass.
I've wanted an interactive map forever.
Now, I just put a prompt into DSH & 43min later, I got my interactive map to show China's new Middle Corridor alternative that bypasses Turkey & goes thru Caucasus & Bulgaria to reach Budapest.
Cost $0.48
I know, youthis w/ your beloved Opus or GPT6.1-Sol, but would it have only cost 48 cent?
That's the power of DeepSeek now. You know it will get better every month. You know DSH keeps getting better. And you know it's basically free, so even El Cheapos like me can afford it.
Basically, the limitation now is fully on me, not on DeepSeek.
I've wanted an interactive map forever.
Now, I just put a prompt into DSH & 43min later, I got my interactive map to show China's new Middle Corridor alternative that bypasses Turkey & goes thru Caucasus & Bulgaria to reach Budapest.
Cost $0.48
I know, youthis w/ your beloved Opus or GPT6.1-Sol, but would it have only cost 48 cent?
That's the power of DeepSeek now. You know it will get better every month. You know DSH keeps getting better. And you know it's basically free, so even El Cheapos like me can afford it.
Basically, the limitation now is fully on me, not on DeepSeek.
19
37
320
23,888
Kimi 3.1 is the best released Chinese model but progression over 3.0 in 3 months is underwhelming.
It’s Opus5.0/GPT-6 Sol level rather than Opus5.5/6.1 level. Not bad, but DeepSeek has clearly progressed faster with its RL than Moonshot. In cyber & some SWE tasks, 1017 is actuDS flash update will catch up to this on SWE & work tasks.
So then what do you do if you are Moonshot? Well, I wonder if it’s just pouring most of its energy on the larger K4. Among top Chinese labs, it seems best at pre training & worst @ post training.
So it’s best bet is keep scaling up. Its models are not efficient like DeepSeek models. It has never pretended otherwise. Even at current px level, it’s overwhelmed by demand so why should it charge less? You should be using DeepSeek if you are just doing regular work stuff.
It’s Opus5.0/GPT-6 Sol level rather than Opus5.5/6.1 level. Not bad, but DeepSeek has clearly progressed faster with its RL than Moonshot. In cyber & some SWE tasks, 1017 is actuDS flash update will catch up to this on SWE & work tasks.
So then what do you do if you are Moonshot? Well, I wonder if it’s just pouring most of its energy on the larger K4. Among top Chinese labs, it seems best at pre training & worst @ post training.
So it’s best bet is keep scaling up. Its models are not efficient like DeepSeek models. It has never pretended otherwise. Even at current px level, it’s overwhelmed by demand so why should it charge less? You should be using DeepSeek if you are just doing regular work stuff.
9
6
123
19,892
China's marine engine export & parts exports. Both of these are rapidly expanding w/ 2026 looking like a banner yr.
Top destination again is Russia. HK is a major intermediary. Other Asian countries are also major customers. Reminder that many Asian countries get parts from Chinw/ some local components & then export to destination.
China's place in the whole Asian export engine cannot be overstated.
Top destination again is Russia. HK is a major intermediary. Other Asian countries are also major customers. Reminder that many Asian countries get parts from Chinw/ some local components & then export to destination.
China's place in the whole Asian export engine cannot be overstated.
2
16
86
4,173
Top global Machine tool exporters in 2026H1 vs 2025 & 2024. Strength of China's machine tool industry against Japan & German is quite evident.
China now exports the most globally, although mostly to global south countries while China itself remains a huge mkt for Japan & German Inc.
Even liled here to both Japan & Germany.
Where would Russian industries be w/o Chinese machines & materials?
China now exports the most globally, although mostly to global south countries while China itself remains a huge mkt for Japan & German Inc.
Even liled here to both Japan & Germany.
Where would Russian industries be w/o Chinese machines & materials?
1
15
99
5,031
Key thing to consider when talking about margin in aI. Yes, when that compute is being used for inference, margin is "high". 85% really isn't high by s/w industry standard.
But you have all this capex that aren't being used for anything. China in fact has this problem w/ a lot of domestic chips that got installed back in 2023-2025 period. ofc, you have a whole bunch of GPUs sitting in warehouses for various American hyperscalers.
I'm not just repeating Ed Zitron actually. I actually have contact w/ ppl involved in AI DC build out.
But you have all this capex that aren't being used for anything. China in fact has this problem w/ a lot of domestic chips that got installed back in 2023-2025 period. ofc, you have a whole bunch of GPUs sitting in warehouses for various American hyperscalers.
I'm not just repeating Ed Zitron actually. I actually have contact w/ ppl involved in AI DC build out.
7
6
72
5,368
DeepSeek had the same problem b4 V4 came out. Chinese AI labs due to compute issues seem to take longer time in pre-training new base model. GLM is going thru that grind now I think.
Can it deliver on its larger base model soon? DS's current cadence improvement is just so fast.
Can it deliver on its larger base model soon? DS's current cadence improvement is just so fast.
国庆前，感觉GLM 5.3 还是国模数一数二的啊。。。
国庆后回来，GLM 5.3 就打不过 Deepseek v4.1 Flash 了。。
现在 GLM 的 token 送都快没人要了。。。
AI 时代的发展真太快了。。。
真不敢想像明年这个时间大模型能力要到哪一步。。
7
4
100
7,640
tphuang retweeted
BREAKING: Kimi K3.1 Pro by @Kimi_Moonshot is projected to rank #1 among open-weights models on DesignArena's Website leaderboard, and is tied for #3 among all models.
Kimi K3.1 Pro has a 15 Elo point lead on their previous model, Kimi K3, on the open weights leaderboard. Among ind only Muse Wonder (max) and Muse Spark 1.3 (xhigh). That ties it with Muse Spark 1.3 Max for third place on the overall Website leaderboard.
Congrats to the @Kimi_Moonshot team on the launch!
Kimi K3.1 Pro has a 15 Elo point lead on their previous model, Kimi K3, on the open weights leaderboard. Among ind only Muse Wonder (max) and Muse Spark 1.3 (xhigh). That ties it with Muse Spark 1.3 Max for third place on the overall Website leaderboard.
Congrats to the @Kimi_Moonshot team on the launch!
20
32
313
14,817