精选头部机构的论文级技术报告、摘要、收录理由与关键要点。 https://polysight.bigmodel-eval.xyz/reports/ zh-CN Google DeepMind 发布 WeatherNext 3：以实时静止卫星马赛克与 FGN mesh transformer 端到端架构，实现每小时、最高 5 公里全球预报，Brightband 独立评测称其迄今最先进准确。 https://polysight.bigmodel-eval.xyz/reports/introducing-weathernext-3-global-weather-ai/ https://polysight.bigmodel-eval.xyz/reports/introducing-weathernext-3-global-weather-ai/ Mon, 19 Oct 2026 00:00:00 GMT Cosmos 3 用双路径 Mixture-of-Transformers 统一语言、图像、视频、音频与动作的理解、生成和控制。 https://polysight.bigmodel-eval.xyz/reports/cosmos-3/ https://polysight.bigmodel-eval.xyz/reports/cosmos-3/ Sat, 17 Oct 2026 00:00:00 GMT DeepSeek-V4 以 CSA/HCA 混合注意力、mHC 与 Muon 支撑百万 token 上下文，并公开预训练、后训练和部署细节。 https://polysight.bigmodel-eval.xyz/reports/deepseek-v4/ https://polysight.bigmodel-eval.xyz/reports/deepseek-v4/ Sat, 17 Oct 2026 00:00:00 GMT DeepSeek-V4.1-Flash 以 CED、CSA2、FP4 KV cache 与 bounded replay 压缩长上下文存储，并加入原生多模态训练。 https://polysight.bigmodel-eval.xyz/reports/deepseek-v4-1-flash/ https://polysight.bigmodel-eval.xyz/reports/deepseek-v4-1-flash/ Sat, 17 Oct 2026 00:00:00 GMT Gemma 4 以多尺寸稠密与 MoE 架构覆盖文本、图像和音频，并围绕推理、长上下文与端侧效率改造模型。 https://polysight.bigmodel-eval.xyz/reports/gemma-4/ https://polysight.bigmodel-eval.xyz/reports/gemma-4/ Sat, 17 Oct 2026 00:00:00 GMT GLM-5 以稀疏注意力、分阶段长上下文训练和异步 Agent RL，面向端到端软件工程与长程工具任务。 https://polysight.bigmodel-eval.xyz/reports/glm-5/ https://polysight.bigmodel-eval.xyz/reports/glm-5/ Sat, 17 Oct 2026 00:00:00 GMT GPT-6 Astra 系统卡详述安全训练、Agent 对齐、监控可见性与达到 Critical 网络安全阈值后的分层防护。 https://polysight.bigmodel-eval.xyz/reports/gpt-6-astra-safety/ https://polysight.bigmodel-eval.xyz/reports/gpt-6-astra-safety/ Sat, 17 Oct 2026 00:00:00 GMT Kimi K3 以混合注意力、稳定稀疏 MoE、原生视觉和百万 token Agent RL，统一长程推理与工具执行。 https://polysight.bigmodel-eval.xyz/reports/kimi-k3/ https://polysight.bigmodel-eval.xyz/reports/kimi-k3/ Sat, 17 Oct 2026 00:00:00 GMT LongCat-Flash-Thinking-2601 以可执行环境扩展、异步强化学习和噪声课程训练提升长程 Agentic 推理及工具使用。 https://polysight.bigmodel-eval.xyz/reports/longcat-flash-thinking-2601/ https://polysight.bigmodel-eval.xyz/reports/longcat-flash-thinking-2601/ Sat, 17 Oct 2026 00:00:00 GMT MiMo-Audio 以高保真音频离散表示和超大规模自回归预训练统一语音理解与生成，并系统评估少样本音频任务迁移。 https://polysight.bigmodel-eval.xyz/reports/mimo-audio/ https://polysight.bigmodel-eval.xyz/reports/mimo-audio/ Sat, 17 Oct 2026 00:00:00 GMT Muse Spark 1.1 评估报告覆盖通用能力、灾难性风险、Agent 安全、对抗鲁棒性与部署缓解措施。 https://polysight.bigmodel-eval.xyz/reports/muse-spark-1-1-evaluation/ https://polysight.bigmodel-eval.xyz/reports/muse-spark-1-1-evaluation/ Sat, 17 Oct 2026 00:00:00 GMT Qwen3.5-Omni 以 Thinker–Talker、混合注意力 MoE 与 ARIA 统一多模态理解、语音生成和实时工具调用。 https://polysight.bigmodel-eval.xyz/reports/qwen3-5-omni/ https://polysight.bigmodel-eval.xyz/reports/qwen3-5-omni/ Sat, 17 Oct 2026 00:00:00 GMT Seed2.0 以 Pro、Lite、Mini 三档模型覆盖语言、多模态与 Agent 任务，并用真实工作流导向的评测框架分析能力边界。 https://polysight.bigmodel-eval.xyz/reports/seed-2-0/ https://polysight.bigmodel-eval.xyz/reports/seed-2-0/ Sat, 17 Oct 2026 00:00:00 GMT Claude Opus 5.5 系统卡集中披露能力、RSP 风险判断、Agent 安全、对齐审计与部署防护。 https://polysight.bigmodel-eval.xyz/reports/claude-opus-5-5-system-card/ https://polysight.bigmodel-eval.xyz/reports/claude-opus-5-5-system-card/ Sat, 17 Oct 2026 00:00:00 GMT