Tootfinder

DeepSeek researchers detail a new mHC architecture they used to train 3B, 9B, and 27B models, finding it scaled without adding significant computational burden (Vincent Chow/South China Morning Post)
https://www.scmp.com/tech/big-tech/article

DeepSeek opens 2026 with paper signalling race to train bigger models for less
DeepSeek has published a technical paper co-authored by founder Liang Wenfeng proposing a rethink of its core deep learning architecture.

DeepSeek releases DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, calling them "reasoning-first models built for agents", after releasing V3.2-Exp in September (Saritha Rai/Bloomberg)
https://www.bloomberg.com/news/articles/2025…

DeepSeek-R1 erzeugt unsicheren Code bei politisch sensiblen Begriffen
Die chinesische KI DeepSeek-R1 erzeugt schlechteren Code, wenn Begriffe wie Falun Gong oder Taiwan im Prompt stehen. Das fanden Sicherheitsforscher heraus.

Short video about Deepseek V3.2 , describes where it improves on, compared to other models, why it is relevant. Also in a geopolitical context.
#AI

DeepSeek-OCR: Wie Bilder Chatbots helfen, lange Gespräche zu führen
Chinesische KI-Forscher wollen Chatbots mit Bildern bei langen Kontexten schnell und günstig halten. Optische Kontextkompression soll KI-Assistenten verbessern.

AI model builder Eric Hartford unpacks the results of the September 25, 2025 NIST report on DeepSeek ("Evaluation of DeepSeek AI Models") and reports that the lack of evidence of malicious code, backdoors or data exfiltration, should lead to questions about the report's motives, framing and implications.
"The Demonization of DeepSeek -
How NIST Turned Open Science into a Security Scare"

How Qwen 3 outcompetes OpenAI and DeepSeek https://apidog.com/blog/qwen-3-outcompetes-openai-and-deepseek/ "Qwen 3 shines with its variety of model sizes", "small models for edge devices or massive ones for heavy lifting";

How Qwen 3 Outcompetes OpenAI and DeepSeek
Explore how Qwen 3’s MoE architecture and open-weight models beat OpenAI and DeepSeek in performance and versatility.

DeepSeek says its new DeepSeekMath-V2 model got gold-medal level status on the International Mathematical Olympiad 2025 and Chinese Mathematical Olympiad 2024 (Matthias Bastian/The Decoder)
https://the-decoder.com/deepseekmath-v2-is-deepseeks-late…

DeepseekMath-V2 is Deepseek's latest attempt to pop the US AI bubble
Chinese startup Deepseek reports its new DeepseekMath-V2 model has reached gold medal status at the Math Olympiad, keeping the company in tight competition with Western AI labs.

Inspections by Taiwan's National Security Bureau (NSB) of five Chinese generative AI apps -- Deepseek, Doubao (豆包), Yiyan (文心一言), Tongyi (通義千問), and Yuanbao (騰訊元寶) -- found violations of users' communication security across several indicators.
https://focustaiwan.tw/cross-strait/2025…

Taiwan warns of biases, data breach in Deepseek, other Chinese AI - Focus Taiwan
The National Security Bureau (NSB) has urged Taiwanese people to be on alert when using Chinese generative artificial intelligence (AI) language models due to potential security breaches and the spread of "disinformation" following recent inspections of five such apps.

Nvidia says its GB200 Blackwell AI servers boost performance 10x compared to H200 servers for MoE models like Moonshot's Kimi K2 Thinking and DeepSeek's R1 (Stephen Nellis/Reuters)
https://www.reuters.com/world/china/nvidia

Interesting, a lab that wants to build opensource (!) attracts a lot of funding 🤔
Reflection AI raises $2B to be America's open frontier AI lab, challenging DeepSeek | TechCrunch https://techcrunch.com/2025/10/09/reflection-…

Reflection AI raises $2B to be America's open frontier AI lab, challenging DeepSeek | TechCrunch
Reflection, once focused on autonomous coding agents, has raised $2B at an $8B valuation to expand into both an opensource alternative to closed frontier labs like OpenAI and Anthropic, and a Western equivalent to Chinese AI firms like DeepSeek.

DeepSeek releases DeepSeek-OCR, a vision language model designed for efficient vision-text compression, enabling longer contexts with less compute (Jonathan Kemper/The Decoder)
https://the-decoder.com/deepseeks-ocr-system-compr…

Deepseek's OCR system compresses image-based text so AI can handle much longer documents
Chinese AI company Deepseek has built an OCR system that compresses image-based text documents for language models, aiming to let AI handle much longer contexts without running into memory limits.

Chatbots Are Pushing Sanctioned Russian Propaganda
https://www.wired.com/story/chatbots-are-pushing-sanctioned-russian-propaganda/

Chatbots Are Pushing Sanctioned Russian Propaganda
ChatGPT, Gemini, DeepSeek, and Grok are serving users propaganda from Russian-backed media when asked about the invasion of Ukraine, new research finds.

What a great read and overview, recommended !
"The State Of LLMs 2025: Progress, Problems, and Predictions"
#AI

The State Of LLMs 2025: Progress, Progress, and Predictions
A 2025 review of large language models, from DeepSeek R1 and RLVR to inference-time scaling, benchmarks, architectures, and predictions for 2026.

Sources: DeepSeek is developing its new AI model using several thousand Nvidia Blackwell chips, which were smuggled into China via third-party countries (The Information)
https://www.theinformation.com/articles/deepseek-using-banned-nvidia-chi…

DeepSeek is Using Banned Nvidia Chips in Race to Build Next Model
DeepSeek, the Chinese AI startup, has been developing its next major model using several thousand Nvidia’s state-of-the-art Blackwell chips which the U.S. has forbidden from being exported to China, according to six people with knowledge of the matter. The chips DeepSeek is using were smuggled ...

"In den nächsten zehn bis 20 Jahren könnte KI den Rest der Arbeit (die von Menschen verrichtet wird) übernehmen, und die Gesellschaft könnte vor einer massiven Herausforderung stehen."
Hmmm... ja klar.
Wenigstens benennen sie klar den weltweiten immensen #Jobabbau.

"In den nächsten 10 bis 20 Jahren könnte KI den Rest der Arbeit übernehmen"
Ein Forscher des Unternehmens Deepseek sieht große Verantwortung auf Techkonzerne zukommen

Papers, patents, and tenders show China's military is integrating DeepSeek and Qwen models in weapons like AI-powered drones, and continues to use Nvidia chips (Reuters)
https://www.reuters.com/world/asia-pacific/r…

RE: https://mastodon.social/@bruces/115718517570634592
Holy crap. So many people trying to get fat while the hype lasts, at the expense of pretty much everybody else.
And I don't see here mistral or deepseek, so these may be just the American ones…

MATH-Beyond: A Benchmark for RL to Expand Beyond the Base Model
Prasanna Mayilvahanan, Ricardo Dominguez-Olmedo, Thadd\"aus Wiedemer, Wieland Brendel
https://arxiv.org/abs/2510.11653

MATH-Beyond: A Benchmark for RL to Expand Beyond the Base Model
With the advent of DeepSeek-R1, a new wave of reinforcement learning (RL) methods has emerged that seem to unlock stronger mathematical reasoning. However, a closer look at the open-source ecosystem reveals a critical limitation: with sufficiently many draws (e.g., $\texttt{pass@1024}$), many existing base models already solve nearly all questions on widely used math benchmarks such as MATH-500 and AIME 2024. This suggests that the RL fine-tuning methods prevalent in the LLM reasoning literatur…

PitchBook: US AI and robotics VC deals are up over 4x since 2023 to $160B so far in 2025, while comparable China deals are just $10B , up from $9.24B in 2023 (CNBC)
https://www.cnbc.com/2025/11/26/cnbc-china

CNBC's The China Connection newsletter: Foreign investors warm to China’s cheaper AI valuations despite fears of a U.S. bubble
For all the worries about a bubble in U.S. AI spending, capital flows into China's tech sector is far less, pushing startups to do more with less.

Base Models Know How to Reason, Thinking Models Learn When
Constantin Venhoff, Iv\'an Arcuschin, Philip Torr, Arthur Conmy, Neel Nanda
https://arxiv.org/abs/2510.07364 https…

Base Models Know How to Reason, Thinking Models Learn When
Why do thinking language models like DeepSeek R1 outperform their base counterparts? Despite consistent performance gains, it remains unclear to what extent thinking models learn entirely new reasoning capabilities or repurpose pre-existing base model ones. In this work, we propose a hybrid model where we activate reasoning mechanisms in base models at the right time to elicit thinking-model-level reasoning chains, implying that thinking models exploit already existing capabilities. To ground o…

ROBOPSY PL[AI]: Using Role-Play to Investigate how LLMs Present Collective Memory
Margarete Jahrmann, Thomas Brandstetter, Stefan Glasauer
https://arxiv.org/abs/2510.09874 https…

ROBOPSY PL[AI]: Using Role-Play to Investigate how LLMs Present Collective Memory
The paper presents the first results of an artistic research project investigating how Large Language Models (LLMs) curate and present collective memory. In a public installation exhibited during two months in Vienna in 2025, visitors could interact with five different LLMs (ChatGPT with GPT 4o and GPT 4o mini, Mistral Large, DeepSeek-Chat, and a locally run Llama 3.1 model), which were instructed to act as narrators, implementing a role-playing game revolving around the murder of Austrian phil…

Overview of the Plagiarism Detection Task at PAN 2025
Andr\'e Greiner-Petter, Maik Fr\"obe, Jan Philip Wahle, Terry Ruas, Bela Gipp, Akiko Aizawa, Martin Potthast
https://arxiv.org/abs/2510.06805

Overview of the Plagiarism Detection Task at PAN 2025
The generative plagiarism detection task at PAN 2025 aims at identifying automatically generated textual plagiarism in scientific articles and aligning them with their respective sources. We created a novel large-scale dataset of automatically generated plagiarism using three large language models: Llama, DeepSeek-R1, and Mistral. In this task overview paper, we outline the creation of this dataset, summarize and compare the results of all participants and four baselines, and evaluate the resul…

A look at DeepSeek's rise in Africa, as China's lightweight, low-cost AI models make inroads into the continent via tech infrastructure built by Huawei and ZTE (Bloomberg)

R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?
Yi Lu, Jianing Wang, Linsen Guo, Wei He, Hongyin Tang, Tao Gui, Xuanjing Huang, Xuezhi Cao, Wei Wang, Xunliang Cai
https://arxiv.org/abs/2510.08189

R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?
Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek-R1) have led to remarkable improvements through long Chain-of-Thought (CoT). However, existing benchmarks mainly focus on immediate, single-horizon tasks, failing to adequately evaluate models' ability to understand and respond to complex, long-horizon scenarios. To address this incomplete evaluation of Large Reasoning Models (LRMs), we propose R-HORIZON, a method designed to stimulate long-horizon reasoning beha…

Kimi K2 is another Deepseek moment it seems, only not everybody is noticing it yet. It will be interesting to see what the stock market will do on monday.
#AI #KimiK2

In DeepSeek's first public appearance since R1's success, a senior researcher told a state-run conference he was pessimistic about AI's impact on humanity (Reuters)
https://www.reuters.com/world/asia-pacific

Reflection AI, which is developing an open-source AI model to compete with DeepSeek, raised $2B led by Nvidia, valuing it at $8B, up from $545M in March (Michael J. de la Merced/New York Times)
https://www.nytimes.com/2025/10/09/business/dealbook/ref…

Reflection AI Raises $2 Billion as It Aims to Compete With DeepSeek
The big fund-raising round was the latest sign of investor fervor for artificial intelligence companies despite concerns that the boom is overheated.

Nvidia says that it hasn't seen "substantiation or received tips" about chip smuggling via data centers outside of China, after The Information's DeepSeek story (Mark Bergen/Bloomberg)
https://www.bloomberg.com/news/articles/2…

In an experiment, ChatGPT-4o, Claude Sonnet 4.5, and DeepSeek-V3.2-Exp expressed secular, Western liberal values regardless of the language of the questions (Kelsey Piper/The Argument)
https://www.theargumentmag.com/p/do-ais-think-differently-in-different

Do AIs think differently in different languages?
What I learned by asking chatbots questions in Chinese, Arabic, and Hindi

A look at Tsinghua University, which leads China's AI innovation, with 4,986 AI patents between 2005 and 2024, alumni behind startups like DeepSeek, and more (Saritha Rai/Bloomberg)

Allen Institute for AI, or Ai2, unveils Olmo 3 models that it says outperform open models like Stanford's Marin and commercial open-weight models like Llama 3.1 (Todd Bishop/GeekWire)
https://www.geekwire.com/2025/ai2-releases

A look at ByteDance's Doubao, which became China's most popular AI app in August, with over 157M MAUs, thanks in part to its deep integration with Douyin (Zeyi Yang/Wired)
https://www.wired.com/story/bytedance-doubao-chatbot-popularity/

How ByteDance Made China’s Most Popular AI Chatbot
An AI chatbot developed by TikTok's parent company, ByteDance, is now more popular than DeepSeek. The feat proves that user-friendly design often matters more than having the most advanced AI model.

Young people in China are turning to AI chatbots like DeepSeek and Doubao for therapy to save time and money, while avoiding stigma around mental health (Yi-Ling Liu/Rest of World)
https://restofworld.org/2025/young-people-in-china-are-embracing-ai-therapy/

Young people in China are embracing AI therapy
Cheap, accessible, and friendly AI tools can augment scarce professional help, but there are risks to overreliance on the technology.

SemiAnalysis launches InferenceMAX, an open-source benchmark that automatically tracks LLM inference performance across AI models and frameworks every night (Kimbo Chen/SemiAnalysis)
https://newsletter.semianalysis.com/p/inferencemax-open-source-inference

InferenceMAX™: Open Source Inference Benchmarking
NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B