Max Lv (@m0d8ye)
@m0d8ye
0 正在关注 0 粉丝
梁总应该没想到这次给老黄带了一波货😇
DeepSeek-V4-Flash-DSpark has now completed the r0b0bench core-subset protocol on 2× NVIDIA GB10. ⚡
Quality
• GSM8K: 95.0%
• HumanEval pass@1: 90.9%
• ARC-Easy: 96.0%
• IFEval lightweight: 79.5%
• BFCL multi-turn: 75.5%
Performance
• Dedicated C1 decode: 79.2 tok/s
• Aggregate output throughput:
C1 86.9 · C2 145.9 · C4 239.3 · C6 337.8 tok/s
• TTFT mean: 221.9 ms
• ITL mean: 68.9 ms
Long context
Needle retrieval passed at 25%, 50%, and 90% depth across a full 1,048,576-token context.
✅ 0 infrastructure errors
✅ Runtime gate passed
✅ 5/5 canaries passed
✅ Package-backed, reproducible result freeze
✅ Valid for publication
These are measured results—not projected capabilities or marketing claims.
显示更多
周末试着让 GPT5.6sol 自动玩了一遍 之前做的 PAL 移植。结果 GPT 一上来就开始改代码作弊跳过各种战斗…全程通关 2 小时 20 分。
欢迎 token 没处烧的朋友来挑战一下。好奇不让AI 作弊，最快多久能通关。
显示更多
把 K3 和一个几乎所有人都无法访问的模型做比较，然后宣称自己胜利✌️
Together with the US Center for AI Standards and Innovation (@NIST), we ran evaluations of Kimi K3 focused on its cyber capabilities. Kimi K3 performs below leading US frontier models on our preliminary cyber evaluations.
显示更多
隔壁组急招人：
JD 看着比较复杂，其实进来以后主要工作就是操控一堆 agent 写各种 GPU kernel，最好有 perf 调优背景。工作地点在上海张江。
觉得 workday 麻烦的可以先发邮件给招聘经理mzhu@nvidia.com，他觉得不错的会联系你填 workday，以免浪费时间。
如果你对自己的简历很有自信也可以直接发给我 mlv@nvidia.com。不过毕竟不是我们组急招，我的 bar 会高一些😇
显示更多
对对对，上海 2006 年内环一万一平
China's Real Estate Market has erased all gains from the last 20 years
不难看出中国成熟制程芯片产已经开始进行全球倾销了。我猜大概率还是和 DRAM/NAND 短缺有关系。需要进一步观察长鑫和长江在低端市场的全球占有率数据。
显示更多
What is happening in Hong Kong?
Chinese semiconductor exports to Hong Kong rose to a record $15 billion in May.
This figure has soared over +200% since early 2025.
At the same time, Chinese chip exports to the world surged to a record $35 billion, tripling over the same period.
As a result, Hong Kong accounted for 43% of China’s semiconductor exports in May.
Over the first 5 months of 2026, China exported $239 billion of chips globally, with Hong Kong absorbing over 50% of this total, a record proportion and up from ~33% a decade ago.
Hong Kong has become a dominant logistics hub for China’s chip trade.
显示更多
所以你是选 xAI 的 Composer 2.5 还是 GitHub 加持的 Kimi 2.7 code? 虽然两者从基础模型来看就是一个东西😂
🆕 @Kimi_Moonshot's Kimi K2.7 Code is now generally available in GitHub Copilot.
This is the first open-weight model offered as a selectable option in the Copilot model picker. 🎉
Early testing shows Kimi K2.7 is a lower-cost option with strong performance comparable to highly popular frontier models.
Try it out now in @code and look out for it in more GitHub Copilot surfaces. 👇
显示更多
有了之前几款 iOS App 的开发经验，这次的 side project 开发要顺利的多。 服务器/客户端：
iOS 客户端：
Testflight：
Android 客户端应该所有人都可以 vibe coding 一个出来，所以就没做了。。。
显示更多
复盘今天市场，小红书指数还是靠谱的。
其实我一直好奇 huggingface 的模型下载为什么不采用 P2P模式，应该可以省下一大笔带宽费用。
当年看了我这篇软文进了我司的同学们现在都已经赚到了第一桶金 所以人生很多时候都是选择大于努力
有句话不知道该不该说：
如果你是一个程序员，
2026年都快过半了， 股票还没赚到钱，
那就别骗自己了
你技术水平真的很拉垮
基于开源大模型和分布式网络的 inference 服务成为可能。接下来就是 web3 inference 服务了吧。
Wow, it has happened!
30.55 tok/s on GLM-5.2 4-bit (from @Zai_org) ran by six RTX Pro 6000's across the USA scattered over WAN!
I can't believe this. It was an insane build, you can read more about it on
显示更多
Google 最先提出了了 Transformer 然后又重新发现了 MoE，并且最早设计了 AI 专用芯片 TPU，为什么现在会落后于两家 startup？
看来 Gemini 真的完了
I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there.
It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together. It has been an honor and a pleasure to work with all of you.
显示更多
