搜索“NVIDIA”

rk Open includes custom GPU kernels optimised for inference through SGLang. Meta said the kernels delivered speedups ranging from 1.03 times to 3.95 times in kernel-level tests on NVIDIA and AMD GPUs. The company also reported a 3% end-to-end performance improvement on AMD MI350 GPUs. The model supports several hardware platforms, including AMD, Apple, Intel and NVIDIA silicon. Developers can further fine-tune it using PyTorch's TorchTitan framework. Meta has released the weights under its Muse

匹配字段正文

els. It runs on inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon, and developers can further tune the model using PyTorch’s TorchTitan. Weights ship in BF16 and NVFP4 versions; running locally requires a GPU with at least 700GB of VRAM for NVFP4 or 2TB for BF16. Meta says Muse Spark Open is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac

匹配字段正文

ls at long-horizon tasks, complex software engineering, and multi-step reasoning.” The two U.S. models Muse Spark Open competes with are essentially Thinking Machines’ Inkling and Nvidia’s Nemotron 3 Ultra. In Meta’s own benchmarks, its new model beats them across the board — often by 20 points or more. Credit: Meta. The closer fight, however, is with Chinese open models that many developers already run today. Spark Open leads Kimi K3, GLM 5.3, and Qwen3.8 on the agentic and tool-use benchmarks,

匹配字段正文

LiLiCorr 投机解码方法在 DGX Spark 上可将词元生成速度最高提升至 2.3 倍。模型还支持 NVFP4 量化，在英伟达 Blackwell 架构上使用 4 位权重和 4 位激活值；在非 Blackwell 平台上，同一检查点使用 4 位权重和 16 位激活值。 英伟达还公布了面向分布式推理的部署优化方案。根据其披露的数据，在使用 16 张 NVIDIA B200 GPU、每名用户每秒生成 50 个词元的测试条件下，NVIDIA Dynamo 配合缓存感知路由、会话亲和性和 KV 缓存卸载，最多可服务 8 倍数量的用户，并将首词元延迟降低 85%。 此次发布还包括基础模型、后训练模型和 NVFP4 量化模型权重，以及训练配方、数据集和强化学习环境。 英伟达表示，Nemotron 3.5 Super VL 在预训练和持续预训练阶段共使用约 26 万亿词元，后训练阶段使用约 380 万个样本，并在数十种环境中进行强化学习。新开放的数据集涵盖智能体工具调用、金融推理、指令遵循、长上下文、公平性和视觉语言理解等领域。 除 Super VL 外，英伟达还推出了两款 Nemotro

匹配字段正文

ad released Gemini 4 Argon globally just days earlier. For robotics, the company had already introduced Gemini Robotics-ER 1.6 , a model for spatial perception and task planning. Training in virtual environments isn't new either: the startup World Labs showed how real scenes can be turned into virtual training worlds, while Nvidia's DreamDojo simulates robot actions as videos. What sets Google's approach apart is the ambition to handle simulation, control, and evaluation all with the same model.

匹配字段正文

foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights. Michal SutterZ.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB RELATED ARTICLESMORE FROM AUTHOR NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model openKylin Outlines Its Vision for Reimagining Linux in the AI Era at Open Source Summit Europe 2026 TypeSafe AI Releases Jev 1.14 With Images, Box Type and 262k Context

匹配字段正文

e.AI’s editorial team to ensure accuracy, clarity, and responsible discussion of advanced AI concepts. Discover More Square Launches Flagship Kiosk With Agentic AI Voice Ordering; NVIDIA Unveils cuPotential as Semiconductor Leaders Adopt CUDA-X; Microsoft Research Publishes AI Weather Forecasting Study in Nature; Google Rolls Out AI Scam Defenses Across Android Messages and Calls; OpenAI Alleges Employee Who Rejoined Apple Took Unreleased Device Files; Baseten Adds Runway’s Gen-4.5 and Aleph 2.0

匹配字段正文

the AI companies to adhere to a slowdown if one of the group does not. Right now, Silicon Valley is split into two camps regarding AI safety risk. On the one hand, Zuckerberg and Nvidia CEO Jensen Huang promote a more accelerationist agenda, arguing that the risks have been overstated and that market forces will push AI companies to self-regulate. On the other hand are more risk-averse leaders, most notably Anthropic CEO Dario Amodei. In September, he published a blog post calling for “speed li

匹配字段正文

ts with emptied context helps them escape “idea basins.” At equal compute it beats single runs and best-of-N on 12 of 13 long-horizon tasks (@DimitrisPapail); Other results:LoGRA (NVIDIA): Full-weight RL that uses less memory than LoRA, 7.2 GiB against 13.2 GiB at 1.5B (@omarsar0); nanoswe: A 1.7B coding agent trained from scratch for about $1,000 of compute scores 25.5% on SWE-bench Verified (@rdolmedo ); Test-writing gap: Frontier agents’ tests catch fewer than 55% of logical mutants (@nielstr

匹配字段正文

n today’s AI rundown: Meta’s Muse Wonder lands at the frontier; Andrej Karpathy sketches AI’s ‘sharp funnel’; Put Meta’s new Muse Wonder to work; Microsoft’s Maia 300 takes aim at Nvidia LATEST DEVELOPMENTS META 🍉 Meta’s Muse Wonder lands at the frontier Image source: Meta The Rundown: Meta Superintelligence Labs just released its long-anticipated Muse Wonder (Watermelon) and Muse Wonder Full, two powerful new AI models, with the full version landing 2 in Artificial Analysis’ intelligence rankin

匹配字段正文

odel that achieves comparable performance to Inkling at a quarter of its size. Inkling-Small is a Mixture-of-Experts transformer with 276B total parameters, 12B active, trained on NVIDIA GB300 NVL72 systems. Like Inkling, it features native reasoning over audio and images, variable thinking effort, a context window of up to 1M tokens, and well-rounded performance across a range of benchmarks. Inkling-Small Inkling Compared to Inkling, Inkling-Small achieves comparable performance with much less

匹配字段正文

raction of compute, with the majority being employed for large-scale RL on synthetic and human-created environments. Inkling was our first major training effort and was trained on NVIDIA GB300 NVL72 systems. Future models will further push the scale of compute across pre-training, post-training and RL. RL at scale We relied on large-scale asynchronous RL to shape model behavior and improve its reasoning and overall performance. The chart below shows the model’s score on a held-out aggregate of r

匹配字段正文

hazardous biosecurity knowledge, by reusing the CBRN pre-training filters from GPT-4o [18]. Our model has a knowledge cutoff of June 2024. Training: The gpt-oss models trained on NVIDIA H100 GPUs using the PyTorch framework [19] with expert-optimized Triton [20] kernels1. The training run for gpt-oss-120b required 2.1 million H100-hours to complete, with gpt-oss-20b needing almost 10x fewer. Both models leverage the Flash Attention [21] algorithms to reduce the memory requirements and accelerat

匹配字段正文

s with emptied context helps them escape “idea basins.” At equal compute it beats single runs and best-of-N on 12 of 13 long-horizon tasks (@DimitrisPapail). Other results: LoGRA (NVIDIA): Full-weight RL that uses less memory than LoRA, 7.2 GiB against 13.2 GiB at 1.5B (@omarsar0). nanoswe: A 1.7B coding agent trained from scratch for about $1,000 of compute scores 25.5% on SWE-bench Verified (@rdolmedo ). Test-writing gap: Frontier agents’ tests catch fewer than 55% of logical mutants (@nielstr

匹配字段正文

raction of compute, with the majority being employed for large-scale RL on synthetic and human-created environments. Inkling was our first major training effort and was trained on NVIDIA GB300 NVL72 systems. Future models will further push the scale of compute across pre-training, post-training and RL. RL at scale We relied on large-scale asynchronous RL to shape model behavior and improve its reasoning and overall performance. The chart below shows the model’s score on a held-out aggregate of r

匹配字段正文

odel that achieves comparable performance to Inkling at a quarter of its size. Inkling-Small is a Mixture-of-Experts transformer with 276B total parameters, 12B active, trained on NVIDIA GB300 NVL72 systems. Like Inkling, it features native reasoning over audio and images, variable thinking effort, a context window of up to 1M tokens, and well-rounded performance across a range of benchmarks. Inkling-Small Inkling Compared to Inkling, Inkling-Small achieves comparable performance with much less

匹配字段正文

complex reasoning AFM 3 Core, AFM 3 Core Advanced, and AFM 3 Cloud, along with ADM 3 Cloud, are all purpose-built for Apple silicon. For AFM 3 Cloud Pro, we worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs in Google Cloud, while maintaining the same guarantees to protect our users’ privacy. More details are available on our Security Research Website. Our third generation of Apple Foundation Models delivers significant advancements across capabilities and quality. In t

匹配字段正文

el works independently in an identical sandbox, with up to 24 hours to profile, rewrite, and benchmark four tasks spanning AttnRes, KDA, and a 512-head-dimension MLA kernel across NVIDIA Hopper GPUs and GPGPU from an alternative vendor. Kimi K3 performed competitively with Fable 5 (with fallback) and substantially outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5. Claude Fable 5 was evaluated by a third party, and its results may include fallback behavior. Across most models, some trajectories inc

匹配字段正文

ing, Michal excels at transforming complex datasets into actionable insights. Michal SutterArchitect Launches Liquid Inference, a Real-Time Auction for LLM Inference; Michal SutterNVIDIA PivotOPD Teaches Multi-Turn AI Agents to Recover From Pivotal Mistakes; Michal SutterAnthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input Tokens; Michal SutterReka Releases Rho-1: A 19B Omni-Reasoning Model That Understands, Generates Video and Outputs Robot Action

匹配字段正文

s files, and checks its own changes. TL;DR Size: 12B total, 2.5B active (64 experts, 8 active), 131,072-token context; Runs on: your own GPUs via vLLM or SGLang (speed tested on 1 NVIDIA H200). GGUF builds start at 7.0 GB for llama.cpp, Ollama and LM Studio; Performance: beats Mellum2 on 15 of 17 listed benchmarks; wins 5 of 17 against Qwen3.5-9B; Best: 82.0 on LiveCodeBench v6, ahead of Qwen3.5-9B (75.4) and Gemma 4 E4B (69.4); Worst: 17.4 on Terminal-Bench 2.1, below Qwen3.5-9B (21.7); Bottom

匹配字段正文