https://devotd.net/en/news/mamba-2-cuda-tile-forward-pass-kernels/devotd.en2026-10-22T12:11:16ZNVIDIA CUDA Tile: Mamba-2 forward-pass acceleration, three-kernel designhttps://devotd.net/ja/news/mamba-2-cuda-tile-forward-pass-kernels/devotd.ja2026-10-22T12:11:16ZNVIDIA CUDA Tile：Mamba-2の順伝播を高速化、3つのカーネル設計https://devotd.net/ko/news/mamba-2-cuda-tile-forward-pass-kernels/devotd.ko2026-10-22T12:11:16ZNVIDIA CUDA Tile: Mamba-2 순방향 가속, 세 커널 설계https://devotd.net/zh-hans/news/mamba-2-cuda-tile-forward-pass-kernels/devotd.zh-cn2026-10-22T12:11:16ZNVIDIA CUDA Tile：Mamba-2 前向加速，三内核设计https://devotd.net/en/news/litellm-1-107-0-supply-chain-incident/devotd.en2026-10-22T11:28:13ZLiteLLM 1.107.0: Compromised artifacts, credential responsehttps://devotd.net/ja/news/litellm-1-107-0-supply-chain-incident/devotd.ja2026-10-22T11:28:13ZLiteLLM 1.107.0：汚染された配布物、認証情報への対応https://devotd.net/ko/news/litellm-1-107-0-supply-chain-incident/devotd.ko2026-10-22T11:28:13ZLiteLLM 1.107.0: 오염 배포물, 자격 증명 대응https://devotd.net/zh-hans/news/litellm-1-107-0-supply-chain-incident/devotd.zh-cn2026-10-22T11:28:13ZLiteLLM 1.107.0：受污染的发行物，凭证应对https://devotd.net/en/news/sage-a15-scientific-agent-model/devotd.en2026-10-22T06:32:24ZSage-A1.5: Scientific agents, orchestration and verificationhttps://devotd.net/ja/news/sage-a15-scientific-agent-model/devotd.ja2026-10-22T06:32:24ZSage-A1.5：科学エージェント、オーケストレーションと検証https://devotd.net/ko/news/sage-a15-scientific-agent-model/devotd.ko2026-10-22T06:32:24ZSage-A1.5: 과학 에이전트, 오케스트레이션·검증https://devotd.net/zh-hans/news/sage-a15-scientific-agent-model/devotd.zh-cn2026-10-22T06:32:24ZSage-A1.5：科学智能体，编排与验证https://devotd.net/en/news/kromoss-moss-image-8-step-partiprompts-benchmarks/devotd.en2026-10-22T01:28:23ZKromoss: 8-step generation, PartiPrompts performancehttps://devotd.net/ja/news/kromoss-moss-image-8-step-partiprompts-benchmarks/devotd.ja2026-10-22T01:28:23ZKromoss：8ステップ生成、PartiPrompts性能https://devotd.net/ko/news/kromoss-moss-image-8-step-partiprompts-benchmarks/devotd.ko2026-10-22T01:28:23ZKromoss: 8스텝 생성, PartiPrompts 성능https://devotd.net/zh-hans/news/kromoss-moss-image-8-step-partiprompts-benchmarks/devotd.zh-cn2026-10-22T01:28:23ZKromoss：8步生成，PartiPrompts 性能https://devotd.net/en/news/torchtitan-update-sota-performance-titanrl/devotd.en2026-10-21T21:40:10ZTorchTitan update: Broader model support, improved large-scale training performancehttps://devotd.net/ja/news/torchtitan-update-sota-performance-titanrl/devotd.ja2026-10-21T21:40:10ZTorchTitan更新：モデル対応の拡大、大規模学習の性能改善https://devotd.net/ko/news/torchtitan-update-sota-performance-titanrl/devotd.ko2026-10-21T21:40:10ZTorchTitan 업데이트: 모델 지원 확대, 대규모 학습 성능 개선https://devotd.net/zh-hans/news/torchtitan-update-sota-performance-titanrl/devotd.zh-cn2026-10-21T21:40:10ZTorchTitan 更新：扩展模型支持，提升大规模训练性能https://devotd.net/en/news/pytorch-flydsl-amd-gpu-kernel-speedups/devotd.en2026-10-21T20:10:46ZPyTorch FlyDSL: Operator support, performance gainshttps://devotd.net/ja/news/pytorch-flydsl-amd-gpu-kernel-speedups/devotd.ja2026-10-21T20:10:46ZPyTorch FlyDSL：演算サポート、性能向上https://devotd.net/ko/news/pytorch-flydsl-amd-gpu-kernel-speedups/devotd.ko2026-10-21T20:10:46ZPyTorch FlyDSL: 연산 지원, 성능 향상https://devotd.net/zh-hans/news/pytorch-flydsl-amd-gpu-kernel-speedups/devotd.zh-cn2026-10-21T20:10:46ZPyTorch FlyDSL：算子支持，性能提升https://devotd.net/en/news/ollama-flint-v0408-rc0/devotd.en2026-10-21T18:45:35ZOllama v0.40.8-rc0: Flint 262K context, macOS supporthttps://devotd.net/ja/news/ollama-flint-v0408-rc0/devotd.ja2026-10-21T18:45:35ZOllama v0.40.8-rc0：Flintの262Kコンテキスト、macOS対応https://devotd.net/ko/news/ollama-flint-v0408-rc0/devotd.ko2026-10-21T18:45:35ZOllama v0.40.8-rc0: Flint 262K 컨텍스트, macOS 지원https://devotd.net/zh-hans/news/ollama-flint-v0408-rc0/devotd.zh-cn2026-10-21T18:45:35ZOllama v0.40.8-rc0：Flint 262K 上下文，macOS 支持https://devotd.net/en/news/nvidia-gtc-berlin-local-ai-inference-acceleration/devotd.en2026-10-21T15:45:54ZNVIDIA GTC Berlin: Local AI devices, faster inferencehttps://devotd.net/ja/news/nvidia-gtc-berlin-local-ai-inference-acceleration/devotd.ja2026-10-21T15:45:54ZNVIDIA GTC Berlin：ローカルAI機器、推論の高速化https://devotd.net/ko/news/nvidia-gtc-berlin-local-ai-inference-acceleration/devotd.ko2026-10-21T15:45:54ZNVIDIA GTC Berlin: 로컬 AI 기기, 추론 가속https://devotd.net/zh-hans/news/nvidia-gtc-berlin-local-ai-inference-acceleration/devotd.zh-cn2026-10-21T15:45:54ZNVIDIA GTC Berlin：本地 AI 设备，推理加速https://devotd.net/en/news/nvidia-rfd4-proteina-reward-guided-protein-design/devotd.en2026-10-21T13:52:04ZRFD4-Proteina: Generation speed, reward-guided searchhttps://devotd.net/ja/news/nvidia-rfd4-proteina-reward-guided-protein-design/devotd.ja2026-10-21T13:52:04ZRFD4-Proteina：生成速度、報酬に基づく探索https://devotd.net/ko/news/nvidia-rfd4-proteina-reward-guided-protein-design/devotd.ko2026-10-21T13:52:04ZRFD4-Proteina: 생성 속도, 보상 기반 탐색https://devotd.net/zh-hans/news/nvidia-rfd4-proteina-reward-guided-protein-design/devotd.zh-cn2026-10-21T13:52:04ZRFD4-Proteina：生成速度，奖励引导搜索https://devotd.net/en/news/qwen-3-8-omni-flash-agentic-audio-visual/devotd.en2026-10-21T09:43:58ZQwen3.8-Omni-Flash: Long-form audio and video understanding, 1M-token contexthttps://devotd.net/ja/news/qwen-3-8-omni-flash-agentic-audio-visual/devotd.ja2026-10-21T09:43:58ZQwen3.8-Omni-Flash：長時間の音声・動画理解、1Mトークンのコンテキストhttps://devotd.net/ko/news/qwen-3-8-omni-flash-agentic-audio-visual/devotd.ko2026-10-21T09:43:58ZQwen3.8-Omni-Flash: 장편 오디오·비디오 이해, 1M 토큰 컨텍스트https://devotd.net/zh-hans/news/qwen-3-8-omni-flash-agentic-audio-visual/devotd.zh-cn2026-10-21T09:43:58ZQwen3.8-Omni-Flash：长音视频理解，1M token 上下文https://devotd.net/en/news/krafton-raon-1-mini-preview/devotd.en2026-10-21T08:35:39ZRaon-1 mini preview: 30B-A4B MoE, up to 2.18× inference throughputhttps://devotd.net/ja/news/krafton-raon-1-mini-preview/devotd.ja2026-10-21T08:35:39ZRaon-1 mini preview：30B-A4B MoE、最大2.18倍の推論スループットhttps://devotd.net/ko/news/krafton-raon-1-mini-preview/devotd.ko2026-10-21T08:35:39ZRaon-1 mini preview: 30B-A4B MoE, 최대 2.18배 추론 처리량https://devotd.net/zh-hans/news/krafton-raon-1-mini-preview/devotd.zh-cn2026-10-21T08:35:39ZRaon-1 mini preview：30B-A4B MoE，最高2.18倍推理吞吐量https://devotd.net/en/news/ollama-v0-40-7/devotd.en2026-10-21T05:24:27ZOllama v0.40.7: Responses API fixes, model metadata exposurehttps://devotd.net/ja/news/ollama-v0-40-7/devotd.ja2026-10-21T05:24:27ZOllama v0.40.7：Responses APIの修正、モデルメタデータの公開https://devotd.net/ko/news/ollama-v0-40-7/devotd.ko2026-10-21T05:24:27ZOllama v0.40.7: Responses API 보정, 모델 메타데이터 노출https://devotd.net/zh-hans/news/ollama-v0-40-7/devotd.zh-cn2026-10-21T05:24:27ZOllama v0.40.7：Responses API 修正、模型元数据展示https://devotd.net/en/news/llama-cpp-b11774-lfm2-vision-layout/devotd.en2026-10-21T03:44:51Zllama.cpp b11774: Aligned LFM2 image processing, size-dependent token and timing changeshttps://devotd.net/ja/news/llama-cpp-b11774-lfm2-vision-layout/devotd.ja2026-10-21T03:44:51Zllama.cpp b11774：LFM2画像処理の整合、サイズ別のトークン数・時間の変化https://devotd.net/ko/news/llama-cpp-b11774-lfm2-vision-layout/devotd.ko2026-10-21T03:44:51Zllama.cpp b11774: LFM2 이미지 처리 정합, 크기별 토큰·시간 변화https://devotd.net/zh-hans/news/llama-cpp-b11774-lfm2-vision-layout/devotd.zh-cn2026-10-21T03:44:51Zllama.cpp b11774：LFM2 图像处理对齐，尺寸引起的令牌与耗时变化https://devotd.net/en/news/halogen-flash-server-0-18-12-npu-image-edit/devotd.en2026-10-21T02:33:50Zhalogen-flash-server 0.18.12: Strix Halo NPU inference acceleration, multi-image editinghttps://devotd.net/ja/news/halogen-flash-server-0-18-12-npu-image-edit/devotd.ja2026-10-21T02:33:50Zhalogen-flash-server 0.18.12：Strix HaloのNPU推論高速化、複数画像の編集https://devotd.net/ko/news/halogen-flash-server-0-18-12-npu-image-edit/devotd.ko2026-10-21T02:33:50Zhalogen-flash-server 0.18.12: Strix Halo NPU 추론 가속, 다중 이미지 편집https://devotd.net/zh-hans/news/halogen-flash-server-0-18-12-npu-image-edit/devotd.zh-cn2026-10-21T02:33:50Zhalogen-flash-server 0.18.12：Strix Halo NPU 推理加速，多图编辑https://devotd.net/en/news/deepspeed-0-19-8/devotd.en2026-10-21T02:06:45ZDeepSpeed v0.19.8: MoE optimization, parallelism expansionhttps://devotd.net/ja/news/deepspeed-0-19-8/devotd.ja2026-10-21T02:06:45ZDeepSpeed v0.19.8：MoEの最適化、並列化の拡張https://devotd.net/ko/news/deepspeed-0-19-8/devotd.ko2026-10-21T02:06:45ZDeepSpeed v0.19.8: MoE 최적화, 병렬화 확장https://devotd.net/zh-hans/news/deepspeed-0-19-8/devotd.zh-cn2026-10-21T02:06:45ZDeepSpeed v0.19.8：MoE 优化，并行化扩展https://devotd.net/en/news/nvidia-openshell-0-1-5/devotd.en2026-10-21T02:05:26ZNVIDIA OpenShell v0.1.5: Policy delivery, audit logginghttps://devotd.net/ja/news/nvidia-openshell-0-1-5/devotd.ja2026-10-21T02:05:26ZNVIDIA OpenShell v0.1.5：ポリシー配信、監査ログhttps://devotd.net/ko/news/nvidia-openshell-0-1-5/devotd.ko2026-10-21T02:05:26ZNVIDIA OpenShell v0.1.5: 정책 전달, 감사 로그https://devotd.net/zh-hans/news/nvidia-openshell-0-1-5/devotd.zh-cn2026-10-21T02:05:26ZNVIDIA OpenShell v0.1.5：策略推送，审计日志https://devotd.net/en/news/harness2-recursive-agent-harness/devotd.en2026-10-20T23:56:26ZHarness²: Automatic harness improvement, higher benchmark scoreshttps://devotd.net/ja/news/harness2-recursive-agent-harness/devotd.ja2026-10-20T23:56:26ZHarness²：ハーネスの自動改善、ベンチマーク性能の向上https://devotd.net/ko/news/harness2-recursive-agent-harness/devotd.ko2026-10-20T23:56:26ZHarness²: 하네스 자동 개선, 벤치마크 성능 상승https://devotd.net/zh-hans/news/harness2-recursive-agent-harness/devotd.zh-cn2026-10-20T23:56:26ZHarness²：自动改进智能体运行框架，提升基准性能