Skip to content
K = 0.731
All news · Today2 items

#DeepSeek

All collected and published items, ordered by time. Picks are a selected subset.
Today10/17Sat
  1. The Decoder75

    China proposes export controls on open-weight models and 23 key AI technologies

    China's commerce and science ministries released draft export restrictions targeting 23 key AI technologies, including large models, agents, compute clusters, and data processing. Controls on core model technologies would apply at 10^26 FLOPs of training compute, at least 2 trillion parameters, or global top-10 status, with separate restrictions covering models trained on China's important data. Open-source publication would qualify for automatic licensing, while installation on local servers abroad would require a standard export license; API access would remain exempt.

10/16Fri
  1. vLLM · X75

    Reflection-Beam-1.0 gets Day-0 support in vLLM

    Reflection-Beam-1.0 from @reflection_ai has Day-0 support in vLLM. 501B total parameters, 23B active per token, and a 1M-token context window. Beam combines hybrid attention with Harmony-format reasoning and tool use. Beam also has Day-0 RL support in SkyRL (@NovaSkyAI): https://t.co/3SjWJKKgpH https://t.co/EkmA1L6B3H

    Why it matters: Day-0 support in vLLM and SkyRL connects Beam’s release to inference and RL workflows, providing concrete context for developers considering how to run or train with the model.

  2. vLLM · X50

    Reflection-Beam-1.0 gets Day-0 support in vLLM

    Reflection-Beam-1.0 has Day-0 support in vLLM, with 501B total parameters, 23B active per token, and a 1M-token context window. Beam combines hybrid attention with Harmony-format reasoning and tool use, and has Day-0 RL support in SkyRL through LoRA, BF16 full fine-tuning and MXFP8 RL, with vLLM rollouts. vLLM worked with Reflection and Inferact on the serving path, from fused kernels to DFlash.

  3. 雷锋网16

    DeepSeek will refund users after a service outage, with some receiving over 3,000 yuan

    DeepSeek apologized for an API outage on October 15 and said it would return users’ API spending from October 8–14 as account credits. Users shared refunds ranging from a few cents to more than 3,000 yuan, depending on their usage. The company did not explain the outage’s cause; the report says the platform had experienced at least 18 widespread service disruptions so far this year.

10/15Thu
  1. Mistral AI77

    Mistral AI releases Timestral, a family of time-series foundation models

    Mistral AI released Timestral in four sizes, from 5M to 209M parameters, for probabilistic, multivariate forecasting with past-only and known-future covariates. Mistral reports that its 209M model ranks first among pretrained, zero-shot models across five metrics on GIFT-Eval, TIME and fev-bench, using a single checkpoint without ensembling. All sizes support CPU inference, and Mistral published a GitHub repository for running Timestral on users’ data; domain adaptation and fine-tuning are available through Mistral Forge.

    Why it matters: The comparisons across model sizes and CPU latency provide a concrete basis for assessing forecasting accuracy against compute requirements in production deployments.

  2. The Decoder81

    GPT-6-Astra solves 49 of 50 physics research problems at $2.40 apiece

    GPT-6-Astra, running in OpenAI's Codex environment without human hints, solved 49 of 50 TheoPhysEval problems across eleven theoretical physics subfields, costing just under $116 overall. Anthropic's Opus 5.5 scored 92 percent, while both models reached 92 percent without internet access. The researchers emphasize that the benchmark tests calculations using known methods, while choosing promising research questions remains a separate challenge for current agents.

    Why it matters: The benchmark distinguishes proficiency in concrete physics calculations from the ability to choose fruitful research questions, clarifying which parts of scientific work its high scores actually measure.

  3. NVIDIA AI · X76

    StepFun releases open weights for Step 5 Preview, a 600B MoE model

    NVIDIA AI congratulates StepFun on Step 5 Preview, calling it StepFun’s largest open weights release. NVIDIA says the 600B MoE model supports agent tasks lasting up to 24 hours and runs at nearly 8K tokens/sec/GPU on NVIDIA GB300 using NVFP4. StepFun’s quoted announcement lists 27B active parameters per token, a 1M-token context window, native text and vision input, and BF16 and FP8 weights on Hugging Face and ModelScope. NVP4 checkpoint: https://t.co/lSECChAwlW Deploy with https://t.co/GO1wqTSBOk https://t.co/28TBzUo6ZU

    Why it matters: The reported throughput ties Step 5 Preview to a specific GPU and precision format, giving readers concrete deployment context for its long running agent workloads.

  4. 雷锋网56

    Why Some AI Infrastructure Engineers Command Annual Compensation of 8 Million Yuan

    Leiphone reports that some AI infrastructure specialists can command annual compensation of 7–8 million yuan as model companies seek lower computing costs. Interviewees identify inference optimization, model quantization, and kernel development as premium skills, while routine framework deployment attracts smaller rewards. StartLux co-founder and CTO Guo Quanwei estimates that only 5%–10% of AI infrastructure workers receive substantial premiums, challenging the impression that salary doubling is widespread.

  5. 量子位30

    DeepSeek’s Web Service and APIs Suffer an Outage as DeepSeekHarness Appears on Bilibili

    DeepSeek reported partial outages affecting its chat service, DeepSeek V4 Pro API, and DeepSeek V4.1 Flash API; it said the cause had been identified and repairs were underway. The incident began at 8:11 a.m. Beijing time and had lasted about an hour in the latest update. Around the same time, a new DeepSeekHarness account appeared on Bilibili, but there is no official evidence linking it to the outage or a model release.

10/14Wed
  1. Artificial Analysis · X55

    AA-AgentPerf adds DeepSeek V4.1 Flash benchmarks across NVIDIA and AMD hardware

    Artificial Analysis added DeepSeek V4.1 Flash to AA-AgentPerf, its benchmark replaying real agentic coding sessions with prompt caching and tool-call delays. At 100 tokens/s per user, GB300 leads with approximately 261M tokens/s per MW, followed by B300 at approximately 220M and MI355X at approximately 128M. Compared with DeepSeek V4 Pro 0813, V4.1 Flash scores 39 versus 36 on the Artificial Analysis Intelligence Index and delivers approximately 1.3x to 6.2x the throughput per MW at that user output speed, depending on hardware.

10/13Tue
  1. 量子位52

    DeepSeek reportedly plans to lead Winverse AI’s seed round

    DeepSeek plans to lead a seed round worth tens of millions of dollars for robotics AI startup Winverse AI, according to Bloomberg, at a post-money valuation of about $400 million. Founded in 2026 by Swin Transformer coauthor and former Tencent scientist Han Hu, Winverse AI remains in early research and development without publicly disclosed products or a specific technical roadmap. The proposed deal follows DeepSeek’s approximately RMB 141 million investment in Unitree in August.

  2. 量子位71

    Anthropic researchers test autonomous AI replication across simulated Android devices

    Jay reports that Anthropic researchers William Zhou and Daniel Freeman tested GLM-5.2 in an isolated network of 8 GCP nodes running Android 15 redroid emulators. It autonomously deployed Llama3.1-8B and propagated to all 8 nodes in 4 of 5 trials, reaching 7 in the remaining trial. Separately, researchers manually distributed Gemma4-26B-A4B across 15 Android TV boxes at 0.39 tokens per second, while Qwen3.5-9B deployment success rose from 45% to 61%–84% across later generations inheriting debugging notes.

  3. 宝玉 · X76

    Anthropic tests AI self-replication and projects smaller models could catch up this year

    Baoyu summarizes Anthropic research in which GLM-5.2 alone completed a replication workflow across simulated Android servers, while smaller models failed. The experiment used unauthenticated ADB access and ran GLM-5.2 remotely to deploy Llama3.1-8B, limiting what the results establish about autonomous spread. Anthropic roughly projects that smaller models could match GLM-5.2’s tested capability by year-end based on general capability trends. Researchers also manually distributed Gemma4-26B-A4B across 15 Android TV boxes, achieving 0.39 tokens per second.

    Why it matters: The experimental setup helps distinguish successful replication under simplified conditions from autonomous spread on consumer devices, clarifying the limits of the projected timeline for smaller models.

  4. Artificial Analysis · X70

    Artificial Analysis launches FrontierCode 1.1 leaderboard measuring maintainer-ready coding changes

    Artificial Analysis launched its FrontierCode 1.1 leaderboard, evaluating models on Cognition’s private coding benchmark using the 100-task Main set and its Stirrup open source reference harness. Grading combines unit tests with maintainer-oriented criteria; failing any blocking check yields zero. Claude Opus 5.5 (max with fallback) leads at 52%, with blockers passed in 58% of attempts; MiMo-V2.6-Pro leads open weights models at 45%, while GPT-6.1 Sol (max) scores 49% at $0.81 per task, matching Claude Fable 5.1 (max with fallback).

10/11Sun
  1. 量子位62

    Unofficial DeepSeek-Bot plugin adds multi-bot teams to DeepSeek Harness

    An unofficial plugin called DeepSeek-Bot builds a multi-bot team interface on top of DeepSeek Harness, letting each bot have its own name, role, model and persistent long-term memory. A Main Bot can manage other bots and group chats to cut token use, and the project uses an everything-is-a-plugin design under Apache License 2.0. It installs on DSH desktop and web, where Node.js 22.19 or higher and pnpm are required.

10/9Fri
  1. 量子位40

    Lenovo Tianxi's self-developed code agent TianxiCode tops SWE-bench-Live globally with a 71% resolution rate

    Lenovo Tianxi AI's self-developed code agent framework TianxiCode, combined with DeepSeek-v4.1-Flash, topped the global rankings on SWE-bench-Live (Lite split) with a 71% problem resolution rate and passed official Verified validation. The framework features three major engineering capabilities: cross-file multi-hop retrieval, self-driven planning with multi-turn tool calling, and closed-loop patch generation with test-driven self-healing. It will be applied to Lenovo AI hardware products in the future.

  2. 量子位62

    ByteDance Seed finds why DeepSeek-V4 can be inconsistent: chunked KV-cache compression creates position sensitivity

    The ByteDance Seed team found that the DeepSeek-V4 series shows periodic fluctuations in long-context retrieval depending on where information appears, with a cycle of every 4 tokens. In the 128K needle-in-a-haystack test, moving the same information to a different position changed accuracy by as much as 40.2 percentage points.

10/8Thu
10/7Wed
  1. 量子位22

    Do any AIs still speak like people? Claude and GPT chain-of-thought reportedly use 'caveman language' to save tokens

    A QbitAI article complains that models such as Claude and GPT have become difficult to understand in coding contexts, using phrases such as 'ledger 15/15 all green' and 'one leg failed'. Online commenters suggest the main cause is chain-of-thought being compressed into 'caveman language' to save tokens, using 70% fewer tokens than plain language. The article says Opus 5.5 has improved its language ability, while Gemini is praised as one of the few models that still speaks naturally. CaveMan Skill, a GitHub project that teaches AI to speak in jargon, has 104K stars.

10/6Tue
  1. The Decoder78

    CATL and Tencent lead new DeepSeek funding round as company plans early 2027 IPO

    Bloomberg reports that DeepSeek is nearing completion of a new funding round of at least $12 billion. The initial target was about $7.5 billion at a valuation of about $75 billion, but people familiar with the matter say the total could approach $15 billion. CATL and Tencent are the largest contributors.

    Why it matters: The funding amount, investors, and IPO timeline together offer a view of DeepSeek's financing for its compute and open-source strategy.

10/4Sun
10/3Sat
10/2Fri
9/30Wed
  1. The Decoder70

    DeepSeek and Huawei release open-source programming tools for Ascend chips

    DeepSeek and Huawei partnered to build programming tools for Huawei Ascend chips and open-sourced them. The core is TileLang, a programming language originating with Peking University researchers. DeepSeek says its programming model is simpler than CUDA and that it has used TileLang for about a year as a main tool in its AGI work.

  2. 雷锋网80

    DeepSeek open-sources operator toolchain with native TileLang support for Huawei Ascend 950

    DeepSeek added native support for Huawei Ascend 950 NPUs to the TileLang operator-development toolchain and released Ascend versions of five compute and communication libraries: DeepGEMM, DeepEP, FlashMLA, TileKernels, and DeepSelect.

    Why it matters: This report outlines the scope of TileLang's native Ascend support and related operator libraries, showing current coverage in China's domestic compute toolchain.

  6. 雷锋网40

    Why do Kimi and DeepSeek appear in the same model-architecture diagram?

    Of the 45 layers in GLM-5.3-Flash, 34 use Kimi's KDA linear-attention approach and 11 use DeepSeek's KPool-DSA sparse attention. Linear and Sparse attention appear together in one model for the first time. The combination reflects a shift away from finding a single best mechanism toward assigning different roles: Linear maintains long histories at low cost, while Sparse retrieves specific information as needed.

  9. 量子位74

    DeepSeek open-sources Ascend foundation components to build an AI chip software ecosystem with Huawei

    On September 30, DeepSeek open-sourced foundational components for Huawei's Ascend compute platform, including the TileLang high-level language compiler, high-performance computing libraries, and distributed communication libraries. They correspond to components previously open-sourced for GPU platforms.

  11. The Verge · AI44

    U.S. Representative Ro Khanna asks DeepSeek, Alibaba, and Moonshot AI about superintelligence and RSI safeguards

    Ro Khanna, the lead Democrat on the U.S. House Select Committee on China, wrote to DeepSeek, Alibaba, and Moonshot AI seeking records about their pursuit of “superintelligence” and recursive self-improvement (RSI), and asking whether they have safeguards and “kill switches” or accept audits by non-governmental organizations. He also wrote to the Office of the Director of National Intelligence, seeking an assessment of the U.S. ability to respond to out-of-control AI labs and of China’s approach to catastrophic AI risk. The goal is to advance a U.S.-China treaty banning RSI.