Tag
该论文系统比较了 2:4 半结构化稀疏化与量化在 LLM 推理中的取舍，发现在仍有更低精度可用时，加入 2:4 稀疏化既更慢又更不准确；只有在达到硬件量化下限（如 FP8/NVFP4）后，2:4 稀疏化才展现出 1.3–1.7× 的线性层加速价值，且需通过重训练恢复质量。
社区用户询问除 Qwen Flash Next 外还有哪些值得尝试的 50-150B 参数量级 MoE 模型，列举了 GPT OSS 120B、Mistral 4 Small 119B、Qwen 3.5 122B A10b 和 Ling 3.0 Flash 等候选，想获取本地部署和非编码用途的推荐。
Bose announced it has acquired Skullcandy, keeping it as a separate brand within its premium consumer audio business to expand the reach of its audio technologies like spatial audio and active noise cancellation to a wider audience.
Unitree unveiled its new H2-X humanoid robot by showcasing it performing powerful kung fu kicks capable of denting car doors, highlighting impressive speed, precision, and force.
Samsung's 55-inch 2026 Frame TV is discounted to its lowest price yet at $897.99 on Amazon, with larger sizes like the 98-inch model also on sale. The art-styled TV features a matte display, motion-sensor art mode, 144Hz refresh rate, and HDR10 Plus support.
WIRED rounds up the best art TV deals for October 2026, highlighting discounts on Samsung's The Frame Pro (LS03HW), the entry-level Frame, and the Hisense S7 CanvasTV, with prices hitting record lows.
一位用户宣布其朋友发布了免费的机器人工程教育平台，包含13种真实开发板的交互式学习内容，涵盖Arduino、树莓派等，旨在替代昂贵的入门套件。
China is doubling its RAM production capacity to stabilize memory chip supply and keep everyday gadget prices from spiking amid rising global demand.
A Linux kernel module hack reportedly unlocks 1.9x faster BF16 tensor math on NVIDIA RTX 3090 cards, boosting BF16 TFLOPS from 82 to 155 and cuBLAS GEMM performance, since NVIDIA ships the cards with fp32-accumulate tensor math at half rate.
At Nvidia's GTC in Berlin, Jensen Huang laid out a European vision centered on modular plug-and-play DSX Pod data centers, physical AI, and AI agents, while announcing robotics platforms and partnerships with VW and Mercedes.
A discussion post asking whether Rust or Python is better for coding with local LLMs, exploring whether Rust's compiler feedback helps compensate for weaker small models running on consumer hardware.
Amazon is launching five new Blink security cameras, including its first 4K model and its farthest-reaching trail camera, all available for preorder in the US.
A tweet compared SKY-001 with GPT-6 in terms of speed and cost on an engine part drawing task—51 times faster and 150 times cheaper—and concluded that this low-cost approach would be widely replicated, while traditional industry remained stuck asking superficial questions.
A hands-on review praising the matte nano-texture inner display on Apple's first-gen iPhone Duo foldable, arguing it sets a new standard for glare reduction and reading comfort that other foldable makers should adopt.
NVIDIA在GTC柏林展示其本地AI硬件全线产品，包括RTX Spark、DGX Spark 64GB和DGX Station，并宣布多项新产品、软件更新及对Nemotron、DeepSeek、Qwen等新模型的首日支持。
NVIDIA推出了装在集装箱里的超级计算机DSX POD，内部布线达17公里，单个容器可提供1GW算力，装机成本约500亿美元，预计年收入达400亿美元。
Unitree introduced the H2-X 'Kung Fu Legend Guardian', a tougher and more durable variant of its H2 humanoid robot boasting around 8 hours of endurance for service and guarding duties.
Menlo Research announces Asimov 1, an open-source humanoid robot built from off-the-shelf parts for agentic robotics, with open CAD, SDK, safety layer, and locomotion policy so agents can understand, repair, and extend the robot.
LEAF 提出首个 LUT 感知的 Transformer/linformer 架构，面向 FPGA 上的超低延迟推理，在喷注标记基准上以更少 LUT 资源实现更高精度，适用于 CERN 等实时系统。
A non-programmer shares how they combined frontier AI for infrastructure with a team of local LLM agents (GPT-OSS 120B on Strix Halo boxes and Gemma 31B on an RTX 5090) to build a massive NAS-based personal library of books, art, photos, and Wikipedia data, using the agents to mine connections and factoids across terabytes of content.