Skip to content
#DeepSeek
TodayNovember 13Friday1 items
  1. Hacker News · AI85

    Forensic analysis links six Chinese frontier models to export-controlled NVIDIA Blackwell GPUs

    Unemployment Labs' forensic analysis of published model weights finds that six models from DeepSeek, Moonshot AI, MiniMax and Ant Group were initialized on NVIDIA Blackwell hardware, while earlier models from the same labs match Hopper. The method recovers each model's random seed and thread layout from preserved weights, and the author says the matches were reproduced on physical H100, B300 and GB200 chips. The report estimates each training run needed roughly 500 to 3,000 Blackwell GPUs over two to three months, and notes the evidence does not establish where the chips were located or whether export controls were violated.

    Why it matters: The report derives hardware evidence solely from published model weights, offering a reproducible method for tracing which GPUs initialized a model without relying on leaks.

November 12Thursday
  1. OpenRouter Announcements28

    How to Benchmark LLMs: A Code-Graded Guide from OpenRouter

    OpenRouter published a guide for benchmarking LLMs on your own application's inputs, grading outputs with code and comparing accuracy, billed cost and request time. On its 60 synthetic tasks, GPT-6.1 Sol passed 180/180 trials across three repeats, while DeepSeek had the lowest response cost at 157/180 passes. Agreement with the public ranking was weak, with a Spearman correlation of 0.109.

  2. vLLM Blog72

    vLLM serves DeepSeek-V4.1-Flash on Vera Rubin NVL72 at 58x GB300 NVL72 throughput

    vLLM reports serving DeepSeek-V4.1-Flash on a single NVIDIA Vera Rubin NVL72 rack at 1.7M tokens/s per GPU with 387 tokens/s per user, 58x the GB300 NVL72, peaking at 2.52M tokens/s and 1,810 tokens/s for a single user. The gains come from kernels rebuilt for Rubin's split memory domains and FP4 tensor cores, a specialized TP8 decode path, and rack-wide KV and Engram pools shared over NVLink. With prefixes pooled across the rack and Dynamo workload-aware routing, the prefix hit rate holds at 96.5-97% even at high concurrency.

    Why it matters: vLLM's third DeepSeek-V4.1-Flash post details how Rubin-specific kernels and rack-wide memory pooling lift throughput 58x over GB300 NVL72, offering a concrete reference for large-scale serving optimization.

November 10Tuesday
  1. Recode China AI38

    Ulanqab, a Remote Chinese City, Becomes an AI Boomtown and a Magnet for Young Workers

    Ulanqab, a once little-known city in Inner Mongolia, has become an emerging AI computing hub as data centers spread across its grasslands, drawing young workers from across China. Kuaishou began operating its first self-built data center there in 2023, and DeepSeek advertised senior operations and maintenance roles paying up to 30,000 yuan ($4,200) a month this April. The labor market is now saturating, with employers raising hiring standards and some considering only bachelor's degree holders.

November 8Sunday
  1. Recode China AI22

    Apple Tests Qwen in Apple Intelligence in China as Tencent Adds Vision to Hy4

    Apple began developer testing of Qwen inside Apple Intelligence in China this week, ahead of a planned December rollout with iOS 27.3, bringing Qwen into Siri, Writing Tools and Visual Intelligence on an opt-in basis. Tencent's November 4 Hy4 release adds visual reasoning and computer use, while Baidu, Moonshot, Ant and Meituan also published or launched new models.

November 6Friday
  1. vLLM Blog77

    REINFORCE Zero: Kernel Alignment Enables Stable Async RL at Policy Lag 64, Extends to DeepSeek-V4.1 and Kimi K3.1

    NVIDIA Research and Inferact's REINFORCE Zero paper shows that aligning rollout and trainer kernels lets unclipped token-level importance sampling correct the remaining weight lag, training stably for 20,000 updates at policy lag 64 and reaching 2.40x the training-step throughput of synchronous training with aligned kernels. With standard kernels, the median rollout-to-learner KL rises 83.33x across the run and online AIME24 pass@1 falls from 34.27% to 13.96%, while a sequence gate stabilizes training but rejects 24.48% of sequences by the last 1,000 updates. The zero-mismatch setup now covers DeepSeek-V4.1-Flash, Kimi K3.1 on release day, and DeepSeek-V4-Flash and Kimi K3.1 on AMD MI355X, with code in Inferact/zero-mismatch-rl and NVIDIA-NeMo/labs-molt.

    Why it matters: NVIDIA Research and Inferact split the async RL rollout-learner gap into weight lag and kernel mismatch, showing that aligning kernels lets unclipped importance sampling hold stable training at policy lag 64.

November 3Tuesday
  1. Recode China AI40

    How Chinese AI Found Its Persona: DeepSeek's Fan-Made Whale Girl and the Shift to Humanized AI Branding

    DeepSeek's fan-created anime mascot "Whale Girl," nicknamed "Big Fat Blue Fish," went viral on Bilibili after creator ZipZipPipe adapted an earlier character using GPT Image 2, with the surge peaking around the DeepSeek-V4.1-Flash release. The character, never officially endorsed by DeepSeek, reflects a broader shift: Chinese AI labs moved from English acronyms like ERNIE and GLM to Chinese names such as Wenxin and Qianwen, then to relatable personas like ByteDance's Doubao.

November 2Monday
  1. Import AI42

    Import AI 479: Spacewar, low-latency tasks versus LLMs, Genie 4

    Import AI 479 covers Realtime-Bench, a 50-task benchmark of 30 web and 20 game tasks built by researchers at UT Austin, Ohio State, UIUC, NVIDIA, Northwestern and Steadyworks to test whether frontier models can act under real-time pressure. GPT-6 Astra (high) scored the highest wall-clock pass rate at 36%, while paused-clock runs let high-reasoning models reach roughly 70% to 80%. The issue also covers the Defense Intelligence Agency's "Challenges to Security in Space" report and Genie 4.

November 1Sunday
  1. Recode China AI22

    CXMT Pushes Into HBM3E, Unitree Q3 Revenue Up 70%, and China's Stock Slide

    CXMT reported Q3 net profit of RMB54.4 billion and revenue of RMB118.71 billion, up 613.2% year on year, while Reuters says it has sent HBM3E samples to domestic chip designers with limited production planned within two months despite yields of just 20–30%. Unitree's first post-listing results showed Q3 revenue up 69.9% to RMB665 million as research spending rose roughly 250%. Moonshot launched Kimi K3.1 Pro with a one-million-token context window, then apologized for rate limits and restored a quarter of subscribers' monthly allowance.

October 25Sunday
  1. Recode China AI38

    DeepSeek's $15B Round, Moonshot's Kimi K3.1 Pro Launch, and China's AI Jobs Policy

    DeepSeek reportedly raised about RMB100 billion at a RMB500 billion pre-money valuation, with Tencent and CATL among the largest backers. Moonshot AI launched Kimi K3.1 Pro with a one-million-token context window and a Highspeed version it says runs two to three times faster, scoring 65.7% on Terminal-Bench 4.0. China's State Council issued guidance with 18 tasks across six areas to pair AI-driven job creation with measures for managing displacement.

October 22Thursday
  1. SemiAnalysis75

    The Huawei-DeepSeek Symbiosis: Ascend 950PR Benchmarks and DeepSeek's Ascend Kernel Support

    SemiAnalysis examines the Huawei-DeepSeek partnership, benchmarking an Ascend 950PR (28 cube cores, 128 GiB HBM) and reviewing DeepSeek's September 30 release of Ascend support for TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. The 950PR reaches 1.6384 TB/s peak HBM bandwidth and beats H100 bandwidth utilization at small message sizes, while its median achieved/attainable compute fraction is 41.5% versus H100's 46.8%; only TileLang is usable so far because the required CANN 9.2.0 was unreleased at writing. The authors argue the software stack is near the frontier even as Huawei hardware trails, and note that most optimizations target supernodes, with vLLM-Ascend still failing out of the box on x86 hosts.

October 18Sunday
  1. Recode China AI40

    China Proposes AI Export Controls, DeepSeek Bets on Robotics, Infinigence Raises $334M

    China's commerce and science ministries proposed adding AI models, agents, chip designs and computing clusters to the technology export-control catalogue, with API, app and SaaS access excluded from the draft rules. DeepSeek reportedly plans to lead a seed round in robotics-AI startup Winverse AI, while Infinigence AI closed a RMB2.3 billion Series C. DeepSeek also released V4.1-Flash-1017 with MIT-licensed weights and a 1M-token context window.

October 13Tuesday
  1. Recode China AI34

    DeepSeek Engineer's Viral Blog on AI Risk and Anthropic Draws Global Media Attention

    A blog by DeepSeek machine learning and kernel engineer Liu Shengyu went viral after he questioned whether frontier AI can now optimize its own low-level GPU code and compared Anthropic CEO Dario Amodei to Adolf Hitler. The post drew coverage from outlets including South China Morning Post and Business Insider, amplifying debate over who should control advanced AI.

May 28, 2025Wednesday