DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and MilesThe SGLang and Miles TeamApril 25, 2026We are thrilled to announce Day-0 support for DeepSeek-V4 across both inference and RL training.
AI系统在各个领域的应用需要明确其行为并评估其可靠性。MLCommons的AI风险与可靠性工作组专注于提高AI可靠性，这不仅能推动市场增长，还能保护社会安全。通过制定详细计划并实施，确保AI系统在不同阶段的可靠性。
AI行业每隔几个月就会推出新一代前沿模型，这些模型的能力不断提升，同时也改变了监管机构、企业和公众需要评估的风险格局。然而，用于衡量这些风险的基准并不会自动更新。本文介绍了MLCommons的AILuminate基准及其Continuous Prompt Stewardship System，该系统通过持续刷新提示数据集，确保基准的长期有效性。系统采用基于心理测量学的指标驱动刷新、闭环数据集再平衡、社区驱动的贡献者模型、双路径审查以及可审计的来源记录等机制，解决基准陈旧和污染问题。AILuminate v1.0包含24,000个人类创作的提示，覆盖12个危害类别，并获得最高长寿分数75分。该系统不仅维护AILuminate的可靠性，还为整个AI基准领域提供宝贵经验，帮助应对基准生命周期挑战。
HiSparse: Turbocharging Sparse Attention with Hierarchical MemoryZhiqiang Xie, Zhangheng Huang, Tingwei HuangApril 10, 2026Why sparse attention leaves performance on the table Self-attention has become a major bottleneck in scaling LLMs to long contexts because of its quadratic compute and memory/IO
MLCommons 近日发布 MLPerf Client v1.6，这是评估个人电脑 AI 性能的最新基准测试套件。该版本针对笔记本电脑、台式机和工作站等设备，模拟真实生成式 AI 任务，如文本摘要、内容创作和代码分析，提供响应速度和吞吐量等标准化指标。新版更新了 Windows ML 和 llama.cpp 等核心运行时，支持 Apple 平台的 MLX with Metal 和 llama.cpp with Metal，提升性能与兼容性。同时，优化了图形界面启动速度、添加进度条，并允许禁用下载确认提示，提高重复测试效率。该基准由 AMD、Intel 等公司协作开发，开源免费，可通过 mlcommons.org/benchmarks/client 下载。（128字）
Highlights of SGLang at NVIDIA GTC 2026The SGLang TeamMarch 31, 2026SGLang came to NVIDIA GTC 2026 with panels, a happy hour, a 200-person meetup, and a hands-on training lab. Three days, five events, one packed week at the center of the LLM ecosystem and left with a lot to share.
MLCommons近日公布了行业标准MLPerf Inference v6.0基准测试套件的最新结果。此次更新包括五个数据中心测试的新增或升级，以及边缘系统的全新物体检测测试。主要亮点有基于GPT-OSS 120B的开源大语言模型基准、扩展的DeepSeek-R1推理测试、首创的顺序推荐基准DLRMv3、文本到视频生成基准，以及基于YOLOv11 Large的物体检测升级。这些变化确保基准测试紧跟AI部署的真实场景，提供全面性能评估。多节点系统提交量激增30%，最大系统达72节点288加速器，彰显行业对大规模推理的重视。24家组织参与，推动AI透明与创新。（128字）
YZ Index testing reveals GPT-4o's code execution (v5) version suffers major performance crisis, with scores dropping from 78.0 to 62.8 out of 100. Six out of seven evaluation dimensions show dramatic declines, raising serious concerns about model stability and reliability.
Elastic EP in SGLang: Achieving Partial Failure Tolerance for DeepSeek MoE DeploymentsThe Mooncake Team, Volcano EngineMarch 25, 20261.
MLPerf Inference v6.0 基准发布显著扩展了对开源大语言模型（LLM）的覆盖。随着行业转向更专业化开源模型，基准测试需适应部署策略与架构变化。本轮引入两大亮点：GPT-OSS 120B 新基准，基于117B参数MoE架构，擅长数学、科学推理与编码；DeepSeek-R1 交互场景，针对实时推理应用，首次标准化 speculative decoding。GPT-OSS 分离性能与准确数据集，准确模式覆盖 AIME 2024（82.92%）、GPQA-Diamond（74.95%）、LiveCodeBench v6（84.68%）；性能模式聚焦 PubMed 摘要生成。DeepSeek-R1 交互场景 TTFT ≤1.5s、TPOT ≤15ms，支持 EAGLE-style decoding。参考实现已在 GitHub 提供。（128字）
GPT-4o suffered a dramatic 10.3-point drop in Material Constraints this week, falling to last place among 11 models, while Baidu's Wenxin 4.0 became the only model to achieve positive growth in core dimensions.
Doubao Pro's stability score crashed from 54.5 to 34.7 in the latest YZ Index evaluation, revealing serious consistency issues that could undermine user trust and enterprise adoption.
ROCm Support for Miles: Large-Scale RL Post-Training on AMD Instinct™ GPUsAMD & Miles TeamMarch 17, 2026Reinforcement learning (RL) has rapidly become a core stage of modern foundation-model development.
Grok 3's stability score crashed from 54.2 to 31.7 points in the latest Winzheng evaluation, exposing a fatal weakness in current AI models that excel at coding but fail at real-world engineering judgment.
GPT-o3's availability score plummeted from 100 to 69 in just one week, exposing fundamental architectural defects rather than isolated issues—a technical accident that reveals systemic imbalances in AI development.
GPT-o3 suffered a catastrophic system failure with stability plummeting from 53 to 28 points and availability dropping from 100 to 69, revealing fundamental architectural flaws rather than typical performance variations.
GPT-o3's long context processing capability collapsed in recent testing, with scores dropping from 62.3 to 28.8 points due to aggressive API rate limiting, exposing serious infrastructure issues at OpenAI.
GPT-4o experiences a catastrophic performance collapse with its usability score plummeting from 100 to 65, caused by overly conservative "strict tool calling" that makes the model refuse to perform basic tasks.
Doubao Pro's stability score plummeted from 54.5 to 34.7 (a 36.3% drop) this week, despite significant improvements in programming and knowledge work dimensions, revealing a concerning pattern of "progress and regression coexisting" that warrants in-depth analysis.
GPT-4o's catastrophic failure in long-context tests, with 5 questions returning rate limit errors, reveals OpenAI's severe infrastructure problems rather than model capability issues.
