Today, leaked notes from a Moonshot investor call and Chinese tech reports laid out their full 2026 and 2027 lineup, and it explains how they plan to fix that problem before building their next giant model.
First, before the end of this year, Moonshot is releasing Kimi K3 Flash. That one is built to replace the older Kimi K2.8 and go head-to-head with DeepSeek V4.1 Flash.
Then around Q1 2027
(right around Lunar New Year),
they are shipping Kimi K3.5. Instead of making the base model bigger, they are keeping it near 3 trillion parameters and training it on 10 trillion post-training tokens. Their whole goal with K3.5 is self-distillation and compression basically teaching the model to reach the same answer using a fraction of the thinking tokens so user quotas actually last.
Once they get those running costs under control, the real monster comes in Q2 or Q3 2027: Kimi K4. According to the leak, Moonshot is scaling Kimi K4 all the way to 5.4 trillion parameters across their H100, H200, and B200 GPU clusters nearly double the size of K3.1 Pro.
That’s probably why Moonshot is looking to rent GPUs in Europe.
Joao Pedro Veiga retweeted
Up to this day people don't trust IQ tests.
How comes we blindly follow benchmarks created by humans themselves, who can't even measure their own intelligence?
Joao Pedro Veiga retweeted
The future is Plazir27.
Union of all tribes.
Collective effort towards pushing the frontier of what's possible.
Understanding that the system we have in place, has no capability to steer the new world order.
Humility, humbleness needs to be prioritized over face value.
AI shouldn't be weaponized. Shouldn't be part of the geopolitical thug of war. It's a desservice to the world. Quite literally, the great equalizer, has arrived.
Which way, western man?
Progress, or the good old neon.
with the usage bar omarchy plugin
XD
insane value tbh / tested on omp for chat models too github.com/can1357/oh-my-pi/…
insane value tbh / tested on omp for chat models too github.com/can1357/oh-my-pi/…
$5.99/month for $165 worth of usage across GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash 1017 & more with Mistral's student plan
Mistral's Education plan is Pro at a student price, with smaller limits.
~$27 of usage for every $1 you pay. Really good value for a student AI plan.
Every month you get:
→ $150 Vibe Code usage (Vibe CLI, VS Code, Vibe web)
→ $15 Studio API usage
→ Vibe Work: chat, web search and image gen on web + mobile
What $165 is worth (one budget, any model):
→ All on GLM 5.3: $165+ (cached tokens $0.14 vs $0.26 at Z․ai)
→ All on GLM 5.3 Flash: $165, same price as Z․ai
→ All on DeepSeek V4.1 Flash 1017: ~$75-135 at DeepSeek's own prices (Mistral never uses its off-peak rate)
How to claim:
1. Sign up at Mistral with your student email
2. On mistral.ai/pricing, click "I'm a student" and check out with a card
3. Copy your Vibe key (the $150) from chat.mistral.ai/code/extensi… → Advanced
4. Create a Studio key (the $15) at console.mistral.ai/api-keys
5. Use both keys in Vibe CLI or OpenCode to spend the full $165 (add the models in the config)
We tested it. The $150 and $15 both showed on our Subscription page, and GLM and DeepSeek ran off both keys. It's legit.
———
Not a student? Mistral Pro is $14.99/month with $25.50 of API credits and about $300 of Vibe Code usage (per users).
Conditions (EDU Plan):
→ Only if you've never used Vibe or Le Chat
→ Student email needed, up to 12 months
→ Taxes may apply depending on your country (we paid $5.99 flat)
→ The $150 isn't on Mistral's pricing page, it shows after you subscribe
→ The $150 Vibe Code usage works in OpenCode. We tested it. It may not be as token-efficient as Vibe CLI.
Joao Pedro Veiga retweeted
4.5k Stars on the repo, much love to everyone!
So much jailbreaking!
All the main AI models and over 20+ smaller LLMs, system prompts and etc!
github.com/Goochbeater/Spiri…
Still works like a charm. Also, kimi k2.8 preview is 50% off.
The meta is using k2.8 preview (charges no cache hits) with swarms of DeepSeek v4.1 flash / glm 5.3-flash
Within kimi code cli, it works wonders. Also, tower mode (new feature, amazing) and swarm mode work like a charm.
Try it out.
For the providers of DeepSeek, I recommend Hyper (20$, daily generous grants to use with any Chinese cheap model), Alibaba Token Plan (70$, no limits, monthly usage only) and also zai for the glm ones.
It currently lacks support for Devin, which will be a game changer if it does.
@Kimi_Moonshot @KimiDevs @KimiAI_Japan
Hell yea. Let us swarm.
DeepSeek V4.1 Pro can beat Opus 5.5.
- DeepSeek V4.1 Pro is no longer just a rumor.
- just leaked into public evaluation scripts.
- Target model on SearchQA
- Paired with V4.1 Flash as worker
- Tested on SkillOpt, MiniAgent (real agent loops)
- RSI highlighted as the main improvement
- Flash already slightly beats Opus 5 on Deep-SWE.
Joao Pedro Veiga retweeted
As a coder for 35+ years, it hurts to say this, but:
the real sloppy code was written by humans, not AI
AI is having a blast hacking that code left and right, right now
People sleeping on the free astra. I went all in on it.
Can i get a medal or something @sama
@OpenAI @OpenAIDevs
I honestly didn't know this was a thing XD
But i expected it.
Actually a badge of honor
Thanks for the usage
Btw, dots is actually an extremely good product. You just gotta exlore how it can continuously provide value.
🔥梁文峰与杨植麟的极致反差!DeepSeek和Kimi的招聘理念,把中国AI的残酷分化说透了!
两家中国大模型公司的招聘标准,反差强烈到近乎对立。
一个赌未来上限,一个要即时战力,背后是赤裸裸的算力鸿沟。
DeepSeek(奥赛原教旨主义)
🔹几乎只看ACM、ICPC、超算等顶级竞赛奖牌,偏爱年轻本科生
🔹不要求丰富工程经验,甚至认为旧的大模型训练经验是包袱
🔹面试重好奇心、技术热情和思维敏捷,而非框架熟练度
🔹接受新人前2-3个月不出成果,用半年到一年把天才养成全栈
Kimi(斯巴达实战派)
🔸哪里缺人就把重任直接扔给你,默认你能扛
🔸不看竞赛光环,死磕是否真碰过AI前沿:模型反作弊、亲手跑大规模实验、跟进新模型发布
🔸面试极难Coding题,考察工程习惯和工具组织能力,写过提效工具直接高分
🔸没耐心等成长,买的是泥泞里打过仗的即战力
底层原因只有一个:算力
DeepSeek算力更充裕,能让奥赛金牌得主花半年读论文、试错底层架构、追求颠覆式创新;
Kimi每张卡都是金贵的,培训新人、排通信Bug是隐性浪费,必须筛出能瞬间把有限算力榨成模型提升的成熟工程师。
两条路都没有绝对对错:
一条是“算力霸权+顶级智力密度”的精英路线,
一条是“极限资源+工程实战”的铁军路线。
在通往AGI的黑夜里,算力是柴火,人才是火种。
柴火够不够,决定你敢不敢养纯粹的天才,还是只能招能立刻上战场的老兵。
#DeepSeek #Kimi #大模型 #AI招聘 #算力 #奥赛 #工程实战 #中国AI #AGI #招聘对比 #梁文峰 #杨植麟
🚨重磅!梁文峰给Agent系统指明方向!DeepSeek 官方给出最新思路:让 AI 从“听话的工具”变成“自己干活的同事”!
以前的 AI Agent 像一个很能干的实习生:你得开一个聊天窗口盯着它,一步步指挥。你同时只能盯住 3~5 个窗口,再多就乱了、忘了上下文。它一等你指令,就停在那儿不动。
DeepSeek Harness 团队提出 Super-agent,想解决这个问题。目标很简单:人只负责定目标、做验收,AI 自己持续干活。
它有四个特点:
🔹连续:不管你在微信群、GitHub 还是邮件里找它,它都记得之前的事,感觉始终是同一个 AI 在跟你说话,不用自己记“哪个窗口说了什么”。
🔹高效:同时推进好几件事,互不耽误,你随时可以插话或改方向。
🔹主动:不用等你 @ 它。群里有人反馈问题、代码评审有评论、邮件来了,它自己观察并决定要不要动手,不需要时就安静待着。
🔹持久:一直在线,电脑关机也不停,出错了能自己恢复。
怎么实现的?核心思路很朴素:
1️⃣一个协调者 + 多个干活的:上面有个“调度员”AI(Router)负责听你说话、分配任务、记住全局;下面多个“工人”AI(Worker)各自独立干活,干完把结果汇报给调度员,再由它统一回复你。你感觉只有一个 AI,背后其实在并行跑多个。
2️⃣事件驱动,而不是等人下命令:群消息、代码评论、CI 结果等,都变成触发信号。系统先用简单规则过滤噪音,再让 AI 判断要不要行动。已经分出去的任务,后续消息直接发给对应的工人,不用每次都绕回调度员。
3️⃣后台先试试再说:看到可能有问题,先悄悄在后台复现、找修复方案。确认有价值再告诉你,没价值就丢掉,不打扰你。
4️⃣人机交互更自然:后台跑长任务时,前台还能继续聊天(不会卡住等结果)。回复风格也按渠道调整:群里短一点,GitHub 上详细一点。
5️⃣一直在线 + 能自愈:计算放云端,真正干活的环境可以留在你自己的电脑上(只跑一个轻量小程序)。外面再加个“监视器”AI,盯着有没有跑偏或卡住,及时纠正。
简单说:Session 不再是你需要操心的东西,变成系统自己的实现细节。人从“全程盯梢”上移到“定方向 + 验收结果”。
他们内部已经在用这套思路跑,公共群里有人反馈问题,即使没 @ AI,它也能自己定位、提修复 PR。
详见原文:mp.weixin.qq.com/s/zoS7GlReW…
#DeepSeek #SuperAgent #AI同事 #Agent #自主AI #Harness #工程实践 #梁文峰 #崔添翼
Joao Pedro Veiga retweeted
聊聊我对 Kimi 和月之暗面的 7️⃣ 点看法。
最近 Kimi 正式发布了 K3.1 Pro,直追 Opus 5.5。参数很大,性能很强。
但它也有争议。一个是额度烧得非常快,一个任务跑下来就见底,今天 Kimi 正式道歉并重置了。
二是宣传时给海外 KOL 的条款被曝光,有些无良自媒体写成、丢中国开源模型的脸。
先说清楚:我并没有接到他们的推广。我只是想站在相对客观的角度,聊聊我自己的看法。
1️⃣ 别太盯着一时的模型强弱,多看公司和团队。
模型这周强一点、下周弱一点,负面舆情今天刷屏、明天没人提。这些都是短暂的事。更值得看的,是背后的人。
月之暗面在科研、审美、融资、商业化,甚至团队氛围上,几乎找不到明显的短板。创始人杨植麟的背景和战略判断力,已经被验证过很多次。
他们就是能一次又一次做出全球最顶级的开源模型。从 K2.5、K2.7,到 K3、K3.1,再往前,Kimi 也是最早把超长上下文做出来的模型。一次是运气。一直这样就是能力。
2️⃣ 模型审美比你想象的重要得多。
我一直觉得,Kimi 和 Anthropic 是现在 AI 圈里最有审美的两个模型团队。
模型能力长期看会趋同。差距只能拉开一小段时间。但审美很难追。早在 Kimi 1.5,甚至更早的时候,它的审美就已经遥遥领先。
你用 AI 做一个很丑的蓝紫色网页,没有人想打开。
Kimi 的艺术审美,最早可以追溯到,杨植麟和周昕宇在清华组过一支叫 Splay 的乐队,杨植麟打鼓兼主唱,周昕宇弹吉他。后来公司名取自 Pink Floyd,Agent 模式叫 OK Computer(Radiohead 的专辑),会员档位用乐谱上的速度记号:andante、moderato、allegretto、allegro。
3️⃣ 快速决策,快速迭代,敢砍沉没成本。
这是我在 Kimi 身上看到的好品质。
很多公司犯了错,钱花了,时间搭进去了,舍不得改,因为有沉默成本。
月之暗面 在 23~24 年登录投流,那时候铺天盖地的广告,直到 2025 年,被一个更强的模型 DeepSeek 一周抢走风头。
他们几乎立刻收缩。杨植麟亲自拍板,砍投放、砍海外、砍低端定制,多余的产品线全砍掉,只做模型。
如果说 AI 时代有什么核心竞争力,快一定是其中之一,我很佩服这种决断能力。
4️⃣ 中国开源模型,也能卖出技术溢价。
过去几年,中国开源模型的打法大多是:性能差不多,价格低很多。说白了就是价格战。
K3 换了个路子。开放权重的同时,把 API 价格放进了全球第一梯队。
这一步挺敢的。等于直接告诉市场:我们的东西值这个价。
能这么定价,前提是模型真的够硬,它也确实做到了。
5️⃣ 有的人天生就适合创业。
月之暗面的股权高度集中,杨植麟是潮汕人,从第一天就握着最终决策权。投资人押的,其实就是他这个人。
好处是快。前面说的砍投放、全面收缩,都是他拍板。代价也很清楚,当初大手笔投流那个错误,同样算在他头上。对和错,都落在一个人身上。
他当年放弃留美,回国创业,原话是:不自己试一次,会后悔一辈子。事实也证明,这种人最契合创业。
5️⃣ 把复杂藏起来,给用户极致的简单。
Kimi 的产品门槛低、上手快、交互顺手,很多技术人看不上,觉得没什么好讲的。
我反而觉得这是 Kimi 的核心竞争力之一。谁能把使用门槛降到最低,把学习时间压到最短,谁就更容易拿到最多的用户。
这份简单也没少干活。背后是更多的研发功课,用户才会觉得顺。所以 Kimi 开发者认,普通用户也接得住。
7️⃣ 人才密度够高,才有资格少管理。
之前有一篇圈内很火的文章,卧底Kim的100个小时,里面揭秘了月之暗面的很多管理细节。
月之暗面几乎没有层级。很少的审批,因为真正想做创造性工作的人才,要的是问题、目标、空间,再加一群足够强的同事。天天被审批和报表追着跑,再强的人也会被磨平。
所以它琢磨的,一直是怎么让最强的那批人尽量少被消耗。
这套别乱学。前提是人才密度先上去。前提不成立,少管理就成了没人管。
额度烧得快、条款有争议,这些该骂就骂,他们也该改。
但如果你问我,未来几年哪一家中国 AI 创业公司最值得长期盯着看,我还是会说,月之暗面。
We're sorry for the rate limits and unnecessary quota consumption on the K3.1 Pro subscription plan. These were our issues, and we should have fixed them sooner.
Today, every subscriber on every plan gets a one-time usage reset, restoring 25% of their monthly quota (usage below 25% resets to zero), equal to one week of usage. Your usage bar will reflect this adjustment by end of day PST.
We've also adjusted the plan's parameters and are reducing unnecessary subagent calls and output length, so your quota goes further.
Thank you for your patience while we put this right.



































