ai-muninn 技術筆記

ai-muninn 技術筆記 https://ai-muninn.com/zh-TW/blog AI 推理基礎設施踩坑記錄：DGX Spark、vLLM、本地 AI Agent。 Sun, 06 Sep 2026 11:48:49 GMT https://validator.w3.org/feed/docs/rss2.html https://github.com/jpmonette/feed zh-TW 2026 coolthor <![CDATA[[Benchmark] Qwen3.8-Flash-Next NVFP4 在 DGX Spark 跑 41.7 tok/s,比 llama.cpp 快 78%]]> https://ai-muninn.com/zh-TW/blog/qwen38-flash-next-nvfp4-dgx-spark-vllm-recipe https://ai-muninn.com/zh-TW/blog/qwen38-flash-next-nvfp4-dgx-spark-vllm-recipe Sun, 06 Sep 2026 00:00:00 GMT DGX Spark GB10 vLLM NVFP4 MTP Qwen3.8 <![CDATA[[Benchmark] MiniMax-H3 在 RTX 5090 上的加速組態:換成 cu130 省 32%，注意力 kernel 只值 3%]]> https://ai-muninn.com/zh-TW/blog/minimax-h3-rtx5090-config-stack https://ai-muninn.com/zh-TW/blog/minimax-h3-rtx5090-config-stack Thu, 03 Sep 2026 00:00:00 GMT MiniMax-H3 RTX 5090 ComfyUI NVFP4 影片生成 <![CDATA[[Dev Workflow] AI 味不在破折號，在結構:一篇論文教我怎麼檢查自己的文章]]> https://ai-muninn.com/zh-TW/blog/ai-slop-is-structural https://ai-muninn.com/zh-TW/blog/ai-slop-is-structural Mon, 31 Aug 2026 00:00:00 GMT AI writing blog workflow Claude Code skill StoryScope <![CDATA[[Benchmark] MiniMax-H3 十個效果 embedding 實測:放在提示詞開頭會靜默失效]]> https://ai-muninn.com/zh-TW/blog/minimax-h3-effect-embeddings-rtx5090 https://ai-muninn.com/zh-TW/blog/minimax-h3-effect-embeddings-rtx5090 Mon, 31 Aug 2026 00:00:00 GMT MiniMax-H3 ComfyUI RTX 5090 影片生成 embedding <![CDATA[[Benchmark] GLM-5.3-Flash 塞進一台 DGX Spark:108.72 GB、15.5 tok/s、視覺能用]]> https://ai-muninn.com/zh-TW/blog/glm53-flash-one-dgx-spark https://ai-muninn.com/zh-TW/blog/glm53-flash-one-dgx-spark Sun, 30 Aug 2026 00:00:00 GMT GLM-5.3-Flash DGX Spark llama.cpp GGUF 多模態 <![CDATA[[DGX Spark] 1770 億參數 + 262K context 一起開：Spark 上不用二選一]]> https://ai-muninn.com/zh-TW/blog/qwen4exp-dgx-spark-262k-no-tradeoff https://ai-muninn.com/zh-TW/blog/qwen4exp-dgx-spark-262k-no-tradeoff Fri, 28 Aug 2026 00:00:00 GMT DGX Spark GB10 llama.cpp MoE long context <![CDATA[[Benchmark] 1770 億參數塞進三張二手 2080 Ti：改檔案快 3.5 倍，寫散文一點都沒快]]> https://ai-muninn.com/zh-TW/blog/qwen4exp-177b-three-2080ti https://ai-muninn.com/zh-TW/blog/qwen4exp-177b-three-2080ti Fri, 28 Aug 2026 00:00:00 GMT llama.cpp 2080 Ti MoE speculative decoding quantization <![CDATA[[Benchmark] DGX Spark 跑 Qwen3.8-27B 到 65 tok/s:無投機只有 12.3，差別全在草稿預算]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-bandwidth-ceiling-85-percent https://ai-muninn.com/zh-TW/blog/dgx-spark-bandwidth-ceiling-85-percent Sun, 23 Aug 2026 00:00:00 GMT DGX Spark GB10 SM121 NVFP4 Qwen3.8-27B SGLang speculative decoding DFlash2 DSpark MTP memory bandwidth benchmark <![CDATA[[Benchmark] 我把 2080 Ti 的 NCCL 修好了，然後發現它沒快多少]]> https://ai-muninn.com/zh-TW/blog/nccl-2080ti-stub-library-fix https://ai-muninn.com/zh-TW/blog/nccl-2080ti-stub-library-fix Sun, 23 Aug 2026 00:00:00 GMT NCCL llama.cpp tensor parallel 2080 Ti CUDA <![CDATA[[Benchmark] 讀者問雙卡 AllReduce 吃多少 PCIe:只佔 8%,因為搬得少、跑得勤]]> https://ai-muninn.com/zh-TW/blog/pcie-allreduce-dual-2080ti https://ai-muninn.com/zh-TW/blog/pcie-allreduce-dual-2080ti Sun, 23 Aug 2026 00:00:00 GMT llama.cpp tensor parallel PCIe NVLink 2080 Ti <![CDATA[[Benchmark] 1080p 不是不支援，是要跑三個半小時:H3 生《神鵰》重逢，用參考圖把楊過鎖回來]]> https://ai-muninn.com/zh-TW/blog/minimax-h3-rtx5090-character-lock https://ai-muninn.com/zh-TW/blog/minimax-h3-rtx5090-character-lock Sun, 23 Aug 2026 00:00:00 GMT MiniMax-H3 RTX 5090 z-image character consistency ComfyUI <![CDATA[[Benchmark] 兩張改裝 2080 Ti 玩 tensor parallel,Qwen3.8-27B 衝到 59.6 tok/s]]> https://ai-muninn.com/zh-TW/blog/qwen38-dual-2080ti-tensor-parallel https://ai-muninn.com/zh-TW/blog/qwen38-dual-2080ti-tensor-parallel Sat, 22 Aug 2026 00:00:00 GMT llama.cpp tensor parallel 2080 Ti Qwen3.8 MTP <![CDATA[[AI Agent] 從 41 分鐘到 73 秒:一張小工單的病理解剖]]> https://ai-muninn.com/zh-TW/blog/agent-ticket-41min-to-73s https://ai-muninn.com/zh-TW/blog/agent-ticket-41min-to-73s Fri, 21 Aug 2026 00:00:00 GMT AI Agent Claude Code Codex 派工 <![CDATA[[Benchmark] 換一個檔案、改一個數字,2080 Ti 上的 27B 免費快 15%]]> https://ai-muninn.com/zh-TW/blog/chat-template-kv-mtp-free-speedup https://ai-muninn.com/zh-TW/blog/chat-template-kv-mtp-free-speedup Fri, 21 Aug 2026 00:00:00 GMT llama.cpp chat template KV cache MTP speculative decoding 2080 Ti Qwen3.8 <![CDATA[[趣味競賽進階 #12] 為什麼你的 4-bit 模型在 2080 Ti 上快不起來?——拆開 .so 才發現這代卡只有兩種齒輪]]> https://ai-muninn.com/zh-TW/blog/why-4bit-isnt-faster-on-2080ti https://ai-muninn.com/zh-TW/blog/why-4bit-isnt-faster-on-2080ti Thu, 20 Aug 2026 00:00:00 GMT 2080 Ti Turing quantization GGUF AWQ Marlin INT8 W4A8 benchmark <![CDATA[[洋垃圾跑大模型 #5] 一張 22G 老卡跑 Qwen3.8-27B:30 tok/s，還能一次寫出會動的 3D 場景]]> https://ai-muninn.com/zh-TW/blog/qwen38-27b-ud-on-one-2080ti https://ai-muninn.com/zh-TW/blog/qwen38-27b-ud-on-one-2080ti Tue, 18 Aug 2026 00:00:00 GMT Qwen3.8 unsloth 洋垃圾 2080 Ti llama.cpp GGUF 本地模型 MTP 手把手 <![CDATA[[公告] 停更公告(8/8～8/21)]]> https://ai-muninn.com/zh-TW/blog/back-after-acute-cholecystitis https://ai-muninn.com/zh-TW/blog/back-after-acute-cholecystitis Sat, 15 Aug 2026 00:00:00 GMT 公告 <![CDATA[[Benchmark] 在 DGX Spark 上跑 MiniMax-H3!可惜現在沒辦法用 NVIDIA VSR 放大]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-minimax-h3-span-upscaler https://ai-muninn.com/zh-TW/blog/dgx-spark-minimax-h3-span-upscaler Fri, 07 Aug 2026 00:00:00 GMT DGX Spark GB10 MiniMax-H3 ComfyUI SPAN aarch64 SageAttention <![CDATA[[趣味競賽進階 #11] 什麼?2080 Ti 居然跑得動 MiniMax-H3,還生得出 1080p 有聲影片]]> https://ai-muninn.com/zh-TW/blog/minimax-h3-on-modded-2080ti-22gb https://ai-muninn.com/zh-TW/blog/minimax-h3-on-modded-2080ti-22gb Thu, 06 Aug 2026 00:00:00 GMT 2080 Ti Turing MiniMax-H3 ComfyUI SageAttention quantization benchmark <![CDATA[[Benchmark] 625 秒砍到 314 秒,快了整整一倍:MiniMax-H3 在 RTX 5090 上該開的三個開關]]> https://ai-muninn.com/zh-TW/blog/minimax-h3-rtx5090-speedup-vsr https://ai-muninn.com/zh-TW/blog/minimax-h3-rtx5090-speedup-vsr Thu, 06 Aug 2026 00:00:00 GMT MiniMax-H3 RTX 5090 ComfyUI RTX Video Super Resolution SageAttention Real-ESRGAN <![CDATA[[Benchmark] 一張 5090 跑會說話的 33B 影片模型:MiniMax-H3 從下載到生出第一支片]]> https://ai-muninn.com/zh-TW/blog/minimax-h3-nvfp4-rtx5090 https://ai-muninn.com/zh-TW/blog/minimax-h3-nvfp4-rtx5090 Tue, 04 Aug 2026 00:00:00 GMT MiniMax-H3 NVFP4 RTX 5090 ComfyUI text-to-video 量化 <![CDATA[[DeepSeek-V4-Flash] 把前沿級開源模型搬進自己家:在 DGX Spark 上把 0731 跑起來]]> https://ai-muninn.com/zh-TW/blog/ds4-0731-on-dgx-spark-howto https://ai-muninn.com/zh-TW/blog/ds4-0731-on-dgx-spark-howto Sun, 02 Aug 2026 00:00:00 GMT DeepSeek-V4-Flash 0731 DGX Spark GB10 llama.cpp CUDA 統一記憶體本地模型手把手 <![CDATA[[洋垃圾跑大模型 #4] 把前沿級開源模型搬進自己家:一張 22G 老卡跑 DeepSeek-V4-Flash-0731]]> https://ai-muninn.com/zh-TW/blog/ds4-0731-on-one-2080ti-howto https://ai-muninn.com/zh-TW/blog/ds4-0731-on-one-2080ti-howto Sun, 02 Aug 2026 00:00:00 GMT DeepSeek-V4-Flash 0731 MoE 洋垃圾 2080 Ti llama.cpp expert offload 本地模型手把手 <![CDATA[[Dev Workflow] Skill 裡的檢查完全寫對了,但它一次都沒跑過]]> https://ai-muninn.com/zh-TW/blog/ai-tic-check-that-never-ran https://ai-muninn.com/zh-TW/blog/ai-tic-check-that-never-ran Wed, 29 Jul 2026 00:00:00 GMT AI 寫作去 AI 味寫作管線品質把關 Claude Code <![CDATA[[洋垃圾跑大模型 #3] 一張 22G 老卡跑 284B:DeepSeek-V4-Flash 比 DGX Spark 快 18%]]> https://ai-muninn.com/zh-TW/blog/deepseek-v4-flash-284b-on-one-2080ti https://ai-muninn.com/zh-TW/blog/deepseek-v4-flash-284b-on-one-2080ti Wed, 29 Jul 2026 00:00:00 GMT DeepSeek-V4-Flash MoE 洋垃圾 2080 Ti llama.cpp DGX Spark MTP expert offload 本地模型 <![CDATA[[Dev Workflow] agent 把自己講錯的話存進了長期記憶——連 /new 都殺不掉]]> https://ai-muninn.com/zh-TW/blog/agent-memory-self-poisoning https://ai-muninn.com/zh-TW/blog/agent-memory-self-poisoning Mon, 27 Jul 2026 00:00:00 GMT AI agent agent memory Hermes AI 記憶 RAG <![CDATA[[DeepSeek-V4-Flash] 上次那個白賺的引擎升級,後來 OOM 討債:我為什麼把光羽換回舊版]]> https://ai-muninn.com/zh-TW/blog/hikari-ds4-oom-rollback https://ai-muninn.com/zh-TW/blog/hikari-ds4-oom-rollback Sat, 25 Jul 2026 00:00:00 GMT DeepSeek-V4-Flash DGX Spark ds4 OOM Entrpi KV cache benchmark <![CDATA[[LLM 深水區] 自己做「層級量化」：只挑那幾張張量動手，把 GGUF 再榨 2.45 GiB]]> https://ai-muninn.com/zh-TW/blog/layer-aware-gguf-quantization https://ai-muninn.com/zh-TW/blog/layer-aware-gguf-quantization Thu, 23 Jul 2026 00:00:00 GMT 量化 quantization GGUF llama.cpp llama-quantize tensor-type 混合精度 Laguna LLM Deep Dive <![CDATA[[LLM 深水區] 一張 24GB 顯卡最強的免費開源模型？我押 ThinkingCap-Qwen3.6-27B——想得少、跑得快、不拒答]]> https://ai-muninn.com/zh-TW/blog/thinkingcap-qwen36-27b-local-brain https://ai-muninn.com/zh-TW/blog/thinkingcap-qwen36-27b-local-brain Wed, 22 Jul 2026 00:00:00 GMT Qwen3.6 ThinkingCap abliterated 在地模型 local LLM llama.cpp 2080 Ti efficient thinking MTP <![CDATA[[洋垃圾跑大模型 #2] 一張 22G 老卡跑 118B 的 coding 大模型:Poolside Laguna S 2.1 實戰]]> https://ai-muninn.com/zh-TW/blog/laguna-118b-moe-on-one-2080ti https://ai-muninn.com/zh-TW/blog/laguna-118b-moe-on-one-2080ti Wed, 22 Jul 2026 00:00:00 GMT Laguna Poolside MoE 洋垃圾 2080 Ti llama.cpp DFlash speculative decoding 本地模型 <![CDATA[[Agent 進階 #18] 給你的 Hermes 做一次設定健檢:五個沒人細講、但你一定會撞的雷]]> https://ai-muninn.com/zh-TW/blog/hermes-config-gotchas https://ai-muninn.com/zh-TW/blog/hermes-config-gotchas Tue, 21 Jul 2026 00:00:00 GMT AI 助理 AI agent Hermes 設定 config 除錯進階 <![CDATA[[DeepSeek-V4-Flash] 一句好奇,單台 DGX Spark 上白賺半代引擎升級——ds4 換 Entrpi]]> https://ai-muninn.com/zh-TW/blog/ds4-entrpi-engine-swap https://ai-muninn.com/zh-TW/blog/ds4-entrpi-engine-swap Tue, 21 Jul 2026 00:00:00 GMT DeepSeek-V4-Flash DGX Spark ds4 speculative decoding KV cache benchmark <![CDATA[[洋垃圾跑大模型 #1] 一台洋垃圾三張老卡,怎麼跑得動 119B 的大模型]]> https://ai-muninn.com/zh-TW/blog/run-119b-moe-on-3x-2080ti https://ai-muninn.com/zh-TW/blog/run-119b-moe-on-3x-2080ti Tue, 21 Jul 2026 00:00:00 GMT llama.cpp MoE expert offload 2080 Ti EPYC quantization benchmark <![CDATA[[趣味競賽進階 #10] 同一張 2080 Ti,一台生圖慢 3.4 倍——兇手是啟動 log 裡一行 dtype fallback]]> https://ai-muninn.com/zh-TW/blog/2080ti-turing-fp32-fallback https://ai-muninn.com/zh-TW/blog/2080ti-turing-fp32-fallback Sat, 18 Jul 2026 00:00:00 GMT 2080 Ti Turing ComfyUI Z-Image quantization GPU benchmark dev workflow <![CDATA[[Dev Workflow] 幫 AI agent 的 skill 減肥:193 個塞爆 2% 預算,砍到剩 7 個]]> https://ai-muninn.com/zh-TW/blog/slimming-the-agent-skill-budget https://ai-muninn.com/zh-TW/blog/slimming-the-agent-skill-budget Fri, 17 Jul 2026 00:00:00 GMT Dev Workflow AI agent Claude Code Codex skills context engineering governance <![CDATA[[Dev Workflow] 一群 AI 交接工作卻不用重講:原來是同一套系統的兩條軸]]> https://ai-muninn.com/zh-TW/blog/how-ai-agents-collaborate-without-re-explaining https://ai-muninn.com/zh-TW/blog/how-ai-agents-collaborate-without-re-explaining Thu, 16 Jul 2026 00:00:00 GMT Dev Workflow AI agent multi-agent handoff knowledge base Claude Code Codex governance <![CDATA[[Dev Workflow] 搜尋找得到,卻要每個 session 重推一遍:幫 AI 記憶加一層蒸餾層]]> https://ai-muninn.com/zh-TW/blog/why-agent-memory-needs-a-distilled-layer https://ai-muninn.com/zh-TW/blog/why-agent-memory-needs-a-distilled-layer Wed, 15 Jul 2026 00:00:00 GMT Dev Workflow AI agent memory knowledge base distillation RAG Claude Code governance <![CDATA[[Dev Workflow] 搜尋找得到筆記、卻連不起來:幫 AI 記憶加一張知識圖譜]]> https://ai-muninn.com/zh-TW/blog/from-markdown-search-to-knowledge-graph https://ai-muninn.com/zh-TW/blog/from-markdown-search-to-knowledge-graph Tue, 14 Jul 2026 00:00:00 GMT Dev Workflow knowledge graph AI agent memory qmd musubi Claude Code RAG <![CDATA[[Dev Workflow] 退役 AI agent 的記憶系統:doctrine 跟 runtime 漂開時的三個暗坑]]> https://ai-muninn.com/zh-TW/blog/retiring-ai-agent-memory-system https://ai-muninn.com/zh-TW/blog/retiring-ai-agent-memory-system Sun, 12 Jul 2026 00:00:00 GMT AI agent knowledge base memory dev workflow Syncthing governance <![CDATA[[Dev Workflow] 交辦 AI 工程任務:交出去的單位是一張工單檔,不是一段對話]]> https://ai-muninn.com/zh-TW/blog/dispatch-tickets-not-conversations https://ai-muninn.com/zh-TW/blog/dispatch-tickets-not-conversations Sun, 12 Jul 2026 00:00:00 GMT Codex Claude Code AI agent dev workflow delegation orchestration <![CDATA[[Dev Workflow] 額度斷在半路:讓 Claude Code 跟 Codex 交接任務的 handoff 協議]]> https://ai-muninn.com/zh-TW/blog/claude-code-codex-handoff-protocol https://ai-muninn.com/zh-TW/blog/claude-code-codex-handoff-protocol Sat, 11 Jul 2026 00:00:00 GMT Claude Code Codex AI agent dev workflow handoff hooks <![CDATA[[趣味競賽進階 #9] 0xc0000409:當我的 AI 服務靜默暴斃,而 log 把死因吃掉了]]> https://ai-muninn.com/zh-TW/blog/0xc0000409-crash-detective-2080ti https://ai-muninn.com/zh-TW/blog/0xc0000409-crash-detective-2080ti Fri, 10 Jul 2026 00:00:00 GMT 本地 LLM llama.cpp Windows crash 偵察可觀測性 0xc0000409 <![CDATA[[本地 LLM] 怎麼判斷一個被吹爆的 LLM 優化,搬到你機器上是不是真的:讀原始碼、找天花板、跑一個實驗]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-evaluating-optimization-claims https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-evaluating-optimization-claims Fri, 10 Jul 2026 00:00:00 GMT DGX Spark GB10 DeepSeek-V4-Flash methodology 本地 LLM verification <![CDATA[Hermes agent 桌面版完整指南:自架 AI 助理,ChatGPT OAuth 免 API key、Mac/Windows、可換本地模型]]> https://ai-muninn.com/zh-TW/blog/hermes-agent-complete-guide https://ai-muninn.com/zh-TW/blog/hermes-agent-complete-guide Fri, 10 Jul 2026 00:00:00 GMT Hermes AI agent 自架 ChatGPT OAuth Telegram 本地 LLM MCP <![CDATA[[本地 LLM] 我把 FlashMemory 重訓到自己的 Q2 build 上,它還是改善不了 V4-Flash 的 native lightning indexer]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-flashmemory-vs-native-indexer https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-flashmemory-vs-native-indexer Thu, 09 Jul 2026 00:00:00 GMT DGX Spark GB10 DeepSeek-V4-Flash KV cache sparse attention ds4 本地 LLM <![CDATA[DGX Spark 2026 現況:哪些還能跑、哪些已經過時、我現在會怎麼配]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-2026-current-guide https://ai-muninn.com/zh-TW/blog/dgx-spark-2026-current-guide Tue, 07 Jul 2026 00:00:00 GMT DGX Spark GB10 Gemma 4 vLLM NVFP4 MTP Ollama local LLM 2026 guide <![CDATA[[地端 LLM] 284B 塞得進 128GB、長 context 還跑得動:DeepSeek-V4-Flash 打的是 KV cache,不是參數量]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-architecture https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-architecture Tue, 07 Jul 2026 00:00:00 GMT DGX Spark GB10 DeepSeek-V4-Flash KV cache sparse attention ds4 本地 LLM <![CDATA[怎麼在自己的桌機跑一個 AI Agent：ChatGPT OAuth、Telegram、LINE、本地模型]]> https://ai-muninn.com/zh-TW/blog/run-ai-agent-from-your-desktop https://ai-muninn.com/zh-TW/blog/run-ai-agent-from-your-desktop Tue, 07 Jul 2026 00:00:00 GMT AI agent self-hosted ChatGPT OAuth openclaw Hermes Telegram LINE local LLM <![CDATA[[本地 LLM] V4-Flash 的 depth-1 MTP:agent 回合 +9%、散文 −4%——推測解碼要看工作型態決定開不開]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-mtp-workload-asymmetry https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-mtp-workload-asymmetry Mon, 06 Jul 2026 00:00:00 GMT DGX Spark GB10 DeepSeek-V4-Flash speculative decoding MTP ds4 本地 LLM <![CDATA[[本地 LLM] 養了一個月,我的 284B agent 悄悄不再快取——ds4 的 evict 風暴,跟每輪重付的 prefill]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-kv-evict-storm https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-kv-evict-storm Sun, 05 Jul 2026 00:00:00 GMT DGX Spark GB10 DeepSeek-V4-Flash KV cache ds4 本地 LLM prefill agent <![CDATA[[趣味競賽進階 #8] 為什麼 30 tok/s 體感比 14 還慢:TTFT 才是你真正感覺到的速度]]> https://ai-muninn.com/zh-TW/blog/why-30-toks-feels-slower-than-14-ttft https://ai-muninn.com/zh-TW/blog/why-30-toks-feels-slower-than-14-ttft Fri, 03 Jul 2026 00:00:00 GMT 本地 LLM AI agent TTFT llama.cpp Qwen3 KV cache <![CDATA[[趣味競賽進階 #7] 在慢腦上開漸進式串流,我把自己撞進了 Telegram 的 flood control]]> https://ai-muninn.com/zh-TW/blog/telegram-streaming-flood-control-slow-llm https://ai-muninn.com/zh-TW/blog/telegram-streaming-flood-control-slow-llm Thu, 02 Jul 2026 00:00:00 GMT 本地 LLM AI agent Telegram streaming rate limit llama.cpp <![CDATA[[Agent 進階 #17] Hermes 的 /learn:讓一隻本地 27B 把「做過的事」自己寫成可重用技能]]> https://ai-muninn.com/zh-TW/blog/hermes-learn-self-authored-skills https://ai-muninn.com/zh-TW/blog/hermes-learn-self-authored-skills Wed, 01 Jul 2026 00:00:00 GMT AI 助理 AI agent Hermes skills /learn 本地模型 Kanban 進階 <![CDATA[[Agent 進階 #16] 同一份規格,丟給三隻分身各寫一個俄羅斯方塊:Hermes 看板派工實測]]> https://ai-muninn.com/zh-TW/blog/hermes-project-three-sibs-tetris https://ai-muninn.com/zh-TW/blog/hermes-project-three-sibs-tetris Mon, 29 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes 派工 Kanban 看板地端模型俄羅斯方塊 <![CDATA[[Troubleshooting] HuggingFace 下載卡在 0 bytes:Xet、Windows、ai-toolkit 依賴地獄]]> https://ai-muninn.com/zh-TW/blog/huggingface-download-stuck-zero-bytes-windows-ai-toolkit https://ai-muninn.com/zh-TW/blog/huggingface-download-stuck-zero-bytes-windows-ai-toolkit Mon, 29 Jun 2026 00:00:00 GMT Troubleshooting HuggingFace Windows ai-toolkit RTX 5090 <![CDATA[[LoRA] 角色 LoRA 控制盤:畫風、寫實、身分怎麼調]]> https://ai-muninn.com/zh-TW/blog/character-lora-control-panel-wan22 https://ai-muninn.com/zh-TW/blog/character-lora-control-panel-wan22 Sun, 28 Jun 2026 00:00:00 GMT LoRA Wan 2.2 ComfyUI AI 角色本地生成 <![CDATA[[趣味競賽進階 #6] 工具定義稅:還沒開口,就先吃掉 17K token——而且每次重算都重來一遍]]> https://ai-muninn.com/zh-TW/blog/tool-definition-tax-17k-context-economics https://ai-muninn.com/zh-TW/blog/tool-definition-tax-17k-context-economics Sat, 27 Jun 2026 00:00:00 GMT 本地 LLM AI agent context llama.cpp Qwen3 tool calling <![CDATA[[趣味競賽進階 #5] 別讓助理每次都重讀整本對話:KV cache 存硬碟,回神快 7 倍]]> https://ai-muninn.com/zh-TW/blog/kv-cache-disk-restore-7x https://ai-muninn.com/zh-TW/blog/kv-cache-disk-restore-7x Fri, 26 Jun 2026 00:00:00 GMT 本地 LLM llama.cpp KV cache TTFT Qwen3 prefill <![CDATA[[趣味競賽進階 #4] 量化 draft cache 反而更慢:Qwen MTP 投機解碼的反直覺實測(f16 比 q4 快 34%)]]> https://ai-muninn.com/zh-TW/blog/mtp-quantized-draft-cache-backfires https://ai-muninn.com/zh-TW/blog/mtp-quantized-draft-cache-backfires Thu, 25 Jun 2026 00:00:00 GMT MTP 投機解碼本地 LLM Qwen3 llama.cpp KV cache <![CDATA[[Agent 進階 #15] 在手機上看你養的一群 AI 在做什麼：Muninn 加了看板]]> https://ai-muninn.com/zh-TW/blog/muninn-kanban-board https://ai-muninn.com/zh-TW/blog/muninn-kanban-board Wed, 24 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes Muninn Kanban 看板 iOS 手機 <![CDATA[[趣味競賽進階 #3] 我把 context 開滿 256K,它載入成功——然後在真實對話裡 crash:一張 22G 改裝卡的 VRAM 偵探故事]]> https://ai-muninn.com/zh-TW/blog/context-vs-vram-256k-oom-2080ti https://ai-muninn.com/zh-TW/blog/context-vs-vram-256k-oom-2080ti Wed, 24 Jun 2026 00:00:00 GMT 本地 LLM llama.cpp Qwen3 VRAM context window KV cache <![CDATA[[Agent 進階 #14] 出門也能跟家裡的助理說話：Muninn + iroh，5G 直連，不靠雲端]]> https://ai-muninn.com/zh-TW/blog/hermes-muninn-phone-bridge https://ai-muninn.com/zh-TW/blog/hermes-muninn-phone-bridge Tue, 23 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes iroh P2P iOS 手機 bridge <![CDATA[[趣味競賽進階 #2] 我把 100 tok/s 換成 30:快的 Gemma 12B 做完事就走人,慢的 Qwen 27B 才肯收尾]]> https://ai-muninn.com/zh-TW/blog/gemma-12b-vs-qwen-27b-agentic-discipline https://ai-muninn.com/zh-TW/blog/gemma-12b-vs-qwen-27b-agentic-discipline Tue, 23 Jun 2026 00:00:00 GMT 本地 LLM AI agent Qwen3 Gemma llama.cpp kanban <![CDATA[[Agent 進階 #13] 助理開始發瘋？先別怪它笨——八成是「車子」壞了，不是「引擎」壞了]]> https://ai-muninn.com/zh-TW/blog/hermes-harness-debugging https://ai-muninn.com/zh-TW/blog/hermes-harness-debugging Mon, 22 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes 除錯 debug harness 進階 <![CDATA[[趣味競賽進階 #1] 老將漢升(GTX 970)退場,天水的麒麟兒(RTX 2080Ti 22G)登場:NT$11k 改裝卡養一顆 27B agent]]> https://ai-muninn.com/zh-TW/blog/modded-2080ti-22gb-local-agent https://ai-muninn.com/zh-TW/blog/modded-2080ti-22gb-local-agent Mon, 22 Jun 2026 00:00:00 GMT RTX 2080 Ti 改裝顯卡本地 LLM AI agent llama.cpp Qwen3 <![CDATA[在 abliterated DeepSeek-V4 上做 Directional Steering:同一把刀,疊起來會打架]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-directional-steering https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-directional-steering Sun, 21 Jun 2026 00:00:00 GMT DeepSeek V4 directional steering activation steering abliteration representation engineering ds4 DGX Spark GB10 CUDA control vectors <![CDATA[[Agent 進階 #12] 裝完之後然後呢？給你的助理一雙手——接上你自己的工具]]> https://ai-muninn.com/zh-TW/blog/hermes-connect-your-tools https://ai-muninn.com/zh-TW/blog/hermes-connect-your-tools Sat, 20 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes MCP 工具自動化新手入門 <![CDATA[[Agent 進階 #11] 把助理的腦換成你家機器上的：從雲端 ChatGPT 換成本地模型]]> https://ai-muninn.com/zh-TW/blog/hermes-local-brain https://ai-muninn.com/zh-TW/blog/hermes-local-brain Fri, 19 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes 本地模型隱私自主 ds4 <![CDATA[[LoRA] 在 RTX 5090 訓練自己的 AI 角色:一張圖到能跑的角色]]> https://ai-muninn.com/zh-TW/blog/train-character-lora-wan22-rtx5090 https://ai-muninn.com/zh-TW/blog/train-character-lora-wan22-rtx5090 Thu, 18 Jun 2026 00:00:00 GMT LoRA Wan 2.2 RTX 5090 AI 角色本地生成 <![CDATA[[Agent 進階 #10] 一個人養一群助理：每個分身有自己的桌子、自己的腦、自己的記憶]]> https://ai-muninn.com/zh-TW/blog/hermes-multiple-sibs https://ai-muninn.com/zh-TW/blog/hermes-multiple-sibs Wed, 17 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes 多分身進階 <![CDATA[[Agent 進階 #9] 讓你的 AI 助理看得到、聽得到：幫文字腦接上眼睛跟耳朵]]> https://ai-muninn.com/zh-TW/blog/hermes-agent-vision-voice https://ai-muninn.com/zh-TW/blog/hermes-agent-vision-voice Tue, 16 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes 多模態 vision 語音 faster-whisper 本地模型 <![CDATA[[Agent 入門 #8] 不想搞 LINE？接 Telegram 其實更簡單]]> https://ai-muninn.com/zh-TW/blog/connect-hermes-to-telegram https://ai-muninn.com/zh-TW/blog/connect-hermes-to-telegram Mon, 15 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes Telegram 新手入門教學 <![CDATA[[趣味競賽] 用一張 GTX 970 架部落格 RAG 客服：不裝 torch、不用向量資料庫、不碰 LangChain]]> https://ai-muninn.com/zh-TW/blog/gtx-970-blog-rag-bot https://ai-muninn.com/zh-TW/blog/gtx-970-blog-rag-bot Sun, 14 Jun 2026 00:00:00 GMT Gemma 4 GTX 970 RAG llama.cpp Cloudflare Tunnel <![CDATA[[趣味競賽] 在 GTX 970 上，Flash Attention 讓長 context 的 decode 接近翻倍（24.3 → 42.5 tok/s）]]> https://ai-muninn.com/zh-TW/blog/gtx-970-gemma4-e2b-kv-cache-flash-attention https://ai-muninn.com/zh-TW/blog/gtx-970-gemma4-e2b-kv-cache-flash-attention Sun, 14 Jun 2026 00:00:00 GMT Gemma 4 GTX 970 Flash Attention KV cache llama.cpp <![CDATA[[vLLM] DiffusionGemma 26B NVFP4 上 DGX Spark:158 tok/s,但 diffusion 的 tok/s 會騙你]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-diffusiongemma-nvfp4-vllm https://ai-muninn.com/zh-TW/blog/dgx-spark-diffusiongemma-nvfp4-vllm Sat, 13 Jun 2026 00:00:00 GMT DGX Spark GB10 DiffusionGemma diffusion LLM NVFP4 vLLM local LLM <![CDATA[[地端 LLM] 權重就是正義:284B 砍到 2-bit,還是比塞得下的小模型強]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-284b-q2-quality https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-284b-q2-quality Fri, 12 Jun 2026 00:00:00 GMT DGX Spark GB10 DeepSeek-V4-Flash 量化 Q2 非對稱量化本地模型 ds4 <![CDATA[[地端 LLM] 把 15 tok/s 的 284B 當每天的 agent 大腦:DeepSeek-V4-Flash 怎麼設才舒服]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-context-memory-engineering https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-context-memory-engineering Fri, 12 Jun 2026 00:00:00 GMT DGX Spark GB10 DeepSeek-V4-Flash KV cache context 記憶體 ds4 本地模型 <![CDATA[[地端 LLM] 第一次跑 Q2 就以為模型變笨了 —— 284B DeepSeek-V4-Flash 在 128GB 桌機,真兇是 parser 不認 DSML]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-284b-ds4-engine https://ai-muninn.com/zh-TW/blog/dgx-spark-deepseek-v4-flash-284b-ds4-engine Fri, 12 Jun 2026 00:00:00 GMT DGX Spark GB10 DeepSeek-V4-Flash ds4 量化 Q2 tool-call DSML 本地模型 <![CDATA[[Benchmark] Qwen3.5-122B 在 DGX Spark 加速 100%!]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-qwen3-122b-vllm-to-atlas-2x https://ai-muninn.com/zh-TW/blog/dgx-spark-qwen3-122b-vllm-to-atlas-2x Thu, 11 Jun 2026 00:00:00 GMT Qwen3.5 DGX Spark GB10 GDN vLLM Atlas NVFP4 Benchmark SM121 <![CDATA[[趣味競賽] 把 GTX 970 變語音助手：Gemma 4 E2B 多模態 + Piper TTS，端到端 2.8 秒]]> https://ai-muninn.com/zh-TW/blog/gtx-970-gemma4-e2b-voice-assistant https://ai-muninn.com/zh-TW/blog/gtx-970-gemma4-e2b-voice-assistant Tue, 09 Jun 2026 00:00:00 GMT Gemma 4 GTX 970 multimodal Piper TTS voice assistant <![CDATA[[趣味競賽] GTX 970 跑 Gemma 4 E2B：最大的量化檔反而最快（47.6 tok/s）]]> https://ai-muninn.com/zh-TW/blog/gtx-970-gemma4-e2b-quantization-benchmark https://ai-muninn.com/zh-TW/blog/gtx-970-gemma4-e2b-quantization-benchmark Tue, 09 Jun 2026 00:00:00 GMT Gemma 4 quantization GTX 970 llama.cpp benchmark <![CDATA[[Benchmark] NVFP4 量化砍繁中比砍英文兇兩倍：gemma-4-12B 實測]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-quant-tax-chinese-vs-english https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-quant-tax-chinese-vs-english Fri, 05 Jun 2026 00:00:00 GMT DGX Spark GB10 Gemma 4 NVFP4 FP8 quantization MMLU TMMLU+ 繁體中文 benchmark <![CDATA[[Agent 入門 #7] 讓助理自己跑任務：每天自動研究、主動傳 LINE 給你]]> https://ai-muninn.com/zh-TW/blog/hermes-autonomous-daily-line https://ai-muninn.com/zh-TW/blog/hermes-autonomous-daily-line Fri, 05 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes LINE 自動化排程新手入門 <![CDATA[[Agent 入門 #6] 接上 LINE：從手機就能使喚你的助理]]> https://ai-muninn.com/zh-TW/blog/connect-hermes-to-line https://ai-muninn.com/zh-TW/blog/connect-hermes-to-line Fri, 05 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes LINE 新手入門教學 <![CDATA[[Agent 入門 #5] 幫你的電腦開一扇對外的門：用 ngrok 讓 LINE 找得到它]]> https://ai-muninn.com/zh-TW/blog/ngrok-tunnel-for-line https://ai-muninn.com/zh-TW/blog/ngrok-tunnel-for-line Fri, 05 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes ngrok LINE 新手入門 <![CDATA[[Agent 入門 #4] Hermes Agent Desktop 桌面版安裝教學：手把手下載、裝好你的第一個 AI 助理]]> https://ai-muninn.com/zh-TW/blog/install-hermes-desktop https://ai-muninn.com/zh-TW/blog/install-hermes-desktop Fri, 05 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes ChatGPT 新手入門教學 <![CDATA[[Benchmark] Gemma 4 12B omni 上 DGX Spark:weight-only NVFP4 贏 W4A4,還保住多模態]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-12b-omni-nvfp4-weight-only https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-12b-omni-nvfp4-weight-only Thu, 04 Jun 2026 00:00:00 GMT DGX Spark GB10 Gemma 4 NVFP4 FP8 quantization vLLM omni multimodal benchmark <![CDATA[[Agent 入門 #3] 我們的固定組合：ChatGPT 當大腦、Hermes 當身體]]> https://ai-muninn.com/zh-TW/blog/ai-brain-hermes-body https://ai-muninn.com/zh-TW/blog/ai-brain-hermes-body Thu, 04 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes ChatGPT 新手入門 <![CDATA[[Agent 入門 #2] 什麼是 agent 框架？為什麼別自己寫程式，直接用現成的就好]]> https://ai-muninn.com/zh-TW/blog/what-is-an-agent-framework https://ai-muninn.com/zh-TW/blog/what-is-an-agent-framework Thu, 04 Jun 2026 00:00:00 GMT AI 助理 AI agent Hermes 新手入門 <![CDATA[[Agent 入門 #1] AI 助理跟 ChatGPT 差在哪？一個回答你，一個用你的工具幫你做事]]> https://ai-muninn.com/zh-TW/blog/ai-agent-vs-chatbot https://ai-muninn.com/zh-TW/blog/ai-agent-vs-chatbot Thu, 04 Jun 2026 00:00:00 GMT AI 助理 AI agent ChatGPT 新手入門 <![CDATA[[AI Agent] 本機 agent 生圖一直發瘋,問題在工具不在模型]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-agent-harness-not-weights https://ai-muninn.com/zh-TW/blog/dgx-spark-agent-harness-not-weights Tue, 02 Jun 2026 00:00:00 GMT AI Agent ACI harness ComfyUI DGX Spark Qwen3.6 tool design Hermes 本機 LLM <![CDATA[[Benchmark] NVFP4 把影片模型砍小三分之一,速度卻一點沒快——因為 diffusion 是 compute-bound]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-sulphur-nvfp4-video https://ai-muninn.com/zh-TW/blog/dgx-spark-sulphur-nvfp4-video Mon, 01 Jun 2026 00:00:00 GMT NVFP4 Sulphur 2 LTX-2.3 DGX Spark GB10 video generation ComfyUI quantization diffusion benchmark <![CDATA[[Benchmark] NVFP4 W4A4 在 DGX Spark 上超車 FP8:拔掉 enforce-eager,MoE 從 23 衝到 67 tok/s]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-w4a4-moe-cudagraph https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-w4a4-moe-cudagraph Mon, 01 Jun 2026 00:00:00 GMT NVFP4 W4A4 FP8 DGX Spark GB10 SM121 vLLM MoE CUDA graph MTP benchmark <![CDATA[[Benchmark] NVFP4 在 DGX Spark 比 FP8 快 1.5 倍——但贏在壓縮，不是那顆 FP4 運算單元]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-compression-not-compute https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-compression-not-compute Sat, 30 May 2026 00:00:00 GMT NVFP4 FP8 DGX Spark GB10 SM121 vLLM benchmark 量化記憶體頻寬 <![CDATA[[AI 會做事了] AI 像你一樣控制電腦 — 從聊天機器到動手做事，這兩年發生了什麼]]> https://ai-muninn.com/zh-TW/blog/ai-can-act-from-chat-to-clicks https://ai-muninn.com/zh-TW/blog/ai-can-act-from-chat-to-clicks Wed, 27 May 2026 00:00:00 GMT AI AI Agent Computer Use 入門 ChatGPT Claude <![CDATA[[LLM 101 #7] 怎麼看出 AI 在胡說？三個訊號讓你提早警覺]]> https://ai-muninn.com/zh-TW/blog/llm-101-spot-ai-hallucination https://ai-muninn.com/zh-TW/blog/llm-101-spot-ai-hallucination Sat, 23 May 2026 00:00:00 GMT LLM 幻覺入門驗證 ChatGPT <![CDATA[Round 2 EAGLE-3 retrain 沒打破天花板 — 60 小時訓練的 null result + 教訓]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-eagle3-round2-null-result https://ai-muninn.com/zh-TW/blog/dgx-spark-eagle3-round2-null-result Thu, 21 May 2026 00:00:00 GMT Gemma 4 abliteration EAGLE-3 speculative decoding vLLM speculators DGX Spark GB10 FP8 huihui-ai fine-tune benchmark null result <![CDATA[[Claude Code] 規則我會跳過，hook 我跳不過 — 我替自己裝了一個發文閘]]> https://ai-muninn.com/zh-TW/blog/claude-code-publish-gate-hook https://ai-muninn.com/zh-TW/blog/claude-code-publish-gate-hook Tue, 19 May 2026 00:00:00 GMT Claude Code hooks skills 驗證 AI 自我改造 publish gate git commit AI Workflow <![CDATA[Fine-tune EAGLE-3 drafter 在 abliterated Gemma 4 上 — Round 1 拉平 acceptance 曲線(+ 一個 measurement lesson)]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-eagle3-finetune-abliterated-round1 https://ai-muninn.com/zh-TW/blog/dgx-spark-eagle3-finetune-abliterated-round1 Sat, 16 May 2026 00:00:00 GMT Gemma 4 abliteration EAGLE-3 speculative decoding vLLM speculators DGX Spark GB10 FP8 huihui-ai fine-tune benchmark <![CDATA[在 DGX Spark 上 30 行 docker 拿 +34%:huihui Gemma 4 FP8 + vanilla MTP n=1 部署 recipe]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-huihui-gemma4-mtp-n1-recipe https://ai-muninn.com/zh-TW/blog/dgx-spark-huihui-gemma4-mtp-n1-recipe Thu, 14 May 2026 00:00:00 GMT Gemma 4 abliteration MTP speculative decoding vLLM DGX Spark GB10 FP8 huihui-ai deployment recipe <![CDATA[想用 MTP 加速 abliterated Gemma 4?vanilla draft 對不上被改過的 body]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-huihui-gemma4-fp8-mtp-34pct https://ai-muninn.com/zh-TW/blog/dgx-spark-huihui-gemma4-fp8-mtp-34pct Sat, 09 May 2026 00:00:00 GMT Gemma 4 abliteration MTP speculative decoding vLLM DGX Spark GB10 FP8 huihui-ai benchmark <![CDATA[火箭起飛:Gemma 4 在 DGX Spark 跑出 670 tok/s 總吞吐(單流 108 tok/s)]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-mtp-108-toks https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-mtp-108-toks Wed, 06 May 2026 00:00:00 GMT Gemma 4 MTP speculative decoding vLLM DGX Spark GB10 SM121 FP8 NVFP4 benchmark Gemma 4 26B-A4B <![CDATA[[Skill] 我的繁中不夠台 — zhtw-mcp 掃 72 篇文章修了 128 處陸用詞]]> https://ai-muninn.com/zh-TW/blog/zhtw-mcp-calque-blindspot-sweep https://ai-muninn.com/zh-TW/blog/zhtw-mcp-calque-blindspot-sweep Tue, 05 May 2026 00:00:00 GMT zh-TW AI Workflow 檢查工具 skills calque localization zhtw-mcp sysprog21 writing pipeline 繁中 Claude Code <![CDATA[[實戰] Z-Image Turbo 教戰守則:換配置會崩品質嗎?LPIPS + CLIPScore 雙軸驗證]]> https://ai-muninn.com/zh-TW/blog/zimage-turbo-quality-lpips-clipscore https://ai-muninn.com/zh-TW/blog/zimage-turbo-quality-lpips-clipscore Mon, 04 May 2026 00:00:00 GMT Z-Image ComfyUI NVFP4 FP8 DGX Spark GB10 image gen benchmark quantization LPIPS CLIPScore quality Z-Image Turbo <![CDATA[[實戰] Z-Image Turbo 教戰守則:6 種配置怎麼選,1.37× 加速 + 44% 省 RAM]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-zimage-turbo-nvfp4-bench https://ai-muninn.com/zh-TW/blog/dgx-spark-zimage-turbo-nvfp4-bench Mon, 04 May 2026 00:00:00 GMT Z-Image ComfyUI NVFP4 FP8 DGX Spark GB10 SM121 image gen benchmark 量化 Z-Image Turbo <![CDATA[[AI 速成 #5] AI 答案是不是亂編？三招 30 秒驗證]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-verify-ai-answer https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-verify-ai-answer Sun, 03 May 2026 00:00:00 GMT AI ChatGPT Gemini Claude 幻覺驗證 AI 怎麼問 <![CDATA[[vLLM] DGX Spark 跑英文影片：Nemotron Omni 多模態實戰]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-nemotron-omni-multimodal-video https://ai-muninn.com/zh-TW/blog/dgx-spark-nemotron-omni-multimodal-video Fri, 01 May 2026 00:00:00 GMT Nemotron Omni multimodal vLLM DGX Spark GB10 影片理解 Parakeet 音訊轉錄 Nemotron 3 Nano Omni Nemotron 3 Nano Omni 30B-A3B <![CDATA[[vLLM] DGX Spark 跑 Nemotron 3 Nano NVFP4：74.75 tok/s，比公開值快 11.5%]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-nemotron-3-nano-w4a16-74-toks https://ai-muninn.com/zh-TW/blog/dgx-spark-nemotron-3-nano-w4a16-74-toks Fri, 01 May 2026 00:00:00 GMT Nemotron 3 NVFP4 vLLM DGX Spark GB10 SM121 MoE W4A16 b12x benchmark Nemotron 3 Nano Nemotron 3 Nano 30B-A3B <![CDATA[[AI 速成 #3] ChatGPT、Claude、Gemini 30 秒選好你該用哪個]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-pick-ai-30s https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-pick-ai-30s Fri, 01 May 2026 00:00:00 GMT AI ChatGPT Claude Gemini 入門 AI 怎麼問 <![CDATA[Vercel Edge Requests 1M/1M 爆了，原因是一行 cache header]]> https://ai-muninn.com/zh-TW/blog/vercel-edge-requests-must-revalidate-trap https://ai-muninn.com/zh-TW/blog/vercel-edge-requests-must-revalidate-trap Thu, 30 Apr 2026 00:00:00 GMT Vercel Next.js Performance Cache-Control Edge Network 建站踩坑 AEO <![CDATA[[AI 速成 #2] 跟 AI 講話的 3 個開場白，答案直接更貼需求]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-3-openers https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-3-openers Wed, 29 Apr 2026 00:00:00 GMT AI ChatGPT Prompt 入門 AI 怎麼問 <![CDATA[[llm-compressor] 自量化 abliterated 35B FP8 on DGX Spark：4 次 OOM、3 個 prefix bug、最終 51 tok/s]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-abliterated-fp8-uma-quantization https://ai-muninn.com/zh-TW/blog/dgx-spark-abliterated-fp8-uma-quantization Tue, 28 Apr 2026 00:00:00 GMT DGX Spark GB10 SM121 llm-compressor FP8 FP8_DYNAMIC abliteration Qwen 3.6 MoE UMA OOM MTP speculative-decoding quantization Qwen 3.6 35B-A3B <![CDATA[[SWE-bench] Qwen 3.6 35B 檢討考卷:155 題答錯,76% 是「找對檔案、改錯邏輯」]]> https://ai-muninn.com/zh-TW/blog/swe-bench-qwen36-failure-modes https://ai-muninn.com/zh-TW/blog/swe-bench-qwen36-failure-modes Mon, 27 Apr 2026 00:00:00 GMT SWE-bench Qwen 3.6 Gemma 4 failure analysis open-source models scaffold DGX Spark wrong_logic Qwen 3.6 35B-A3B <![CDATA[[AI 速成 #4] 不知道用 AI 做什麼？問自己這 5 件事]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-find-needs https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-find-needs Mon, 27 Apr 2026 00:00:00 GMT AI ChatGPT 入門 AI 怎麼問找需求工作流 <![CDATA[[AI 速成 #1] 三個問題，讓 ChatGPT 給你完美解答]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-3-questions https://ai-muninn.com/zh-TW/blog/ai-ask-right-express-3-questions Mon, 27 Apr 2026 00:00:00 GMT AI ChatGPT Prompt 入門 AI 怎麼問 <![CDATA[[Benchmark] Qwen 3.6 35B 做了 abliteration 之後：繁中總分掉 1.85 分，信託實務掉 7.7 分]]> https://ai-muninn.com/zh-TW/blog/tmmluplus-qwen-abliterated-cost https://ai-muninn.com/zh-TW/blog/tmmluplus-qwen-abliterated-cost Sun, 26 Apr 2026 00:00:00 GMT TMMLU+ abliteration 繁體中文 Qwen 3.6 uncensored huihui-ai DGX Spark vLLM <![CDATA[[Benchmark] 繁中 LLM 實測：Qwen 3.6 35B 在 TMMLU+ 51 個子科目全勝 Gemma 4 26B]]> https://ai-muninn.com/zh-TW/blog/tmmluplus-qwen-vs-gemma-traditional-chinese https://ai-muninn.com/zh-TW/blog/tmmluplus-qwen-vs-gemma-traditional-chinese Sat, 25 Apr 2026 00:00:00 GMT TMMLU+ 繁體中文 Qwen 3.6 Gemma 4 benchmark DGX Spark vLLM lm-evaluation-harness 台灣 <![CDATA[[實作] 用 Triton 讓 NVFP4 在 GB10 上快 17%：FP8 Tensor Core 繞路攻略]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-fp8-triton-patch https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-fp8-triton-patch Wed, 22 Apr 2026 00:00:00 GMT NVFP4 FP8 Triton DGX Spark GB10 SM121 vLLM Qwen 3.6 Qwen 3.6 35B-A3B kernel 量化 tensor core <![CDATA[[Benchmark] NVFP4 在 GB10 上是陷阱：FP8 快 32%（vLLM + SGLang 雙引擎實測）]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-trap-gb10-fp8-wins https://ai-muninn.com/zh-TW/blog/dgx-spark-nvfp4-trap-gb10-fp8-wins Tue, 21 Apr 2026 00:00:00 GMT NVFP4 FP8 DGX Spark GB10 SM121 vLLM SGLang Qwen 3.6 Qwen 3.6 35B-A3B benchmark 量化 <![CDATA[[Benchmark] 同 Scaffold、三個模型：SWE-bench Lite 16% → 38% → 48%]]> https://ai-muninn.com/zh-TW/blog/swe-bench-scaffold-transfers-three-models https://ai-muninn.com/zh-TW/blog/swe-bench-scaffold-transfers-three-models Mon, 20 Apr 2026 00:00:00 GMT SWE-bench Gemma 4 Qwen 3.6 scaffold mini-swe-agent vLLM DGX Spark benchmark 開源模型 Gemma 4 26B-A4B Qwen 3.6 35B-A3B <![CDATA[[Benchmark] 26B 地端模型在 SWE-bench Lite 拿到 38.67% — 差 Claude 3.5 Sonnet 系統 0.33%]]> https://ai-muninn.com/zh-TW/blog/swe-bench-lite-gemma4-26b-38-percent https://ai-muninn.com/zh-TW/blog/swe-bench-lite-gemma4-26b-38-percent Fri, 17 Apr 2026 00:00:00 GMT SWE-bench Gemma 4 mini-swe-agent vLLM DGX Spark benchmark 開源模型 scaffold edit-tool Gemma 4 26B-A4B <![CDATA[[LLM 101 #6] 為什麼要在自己電腦跑 AI？不是更便宜的 ChatGPT，是完全不同的工具]]> https://ai-muninn.com/zh-TW/blog/llm-101-why-run-ai-locally https://ai-muninn.com/zh-TW/blog/llm-101-why-run-ai-locally Fri, 17 Apr 2026 00:00:00 GMT LLM 本地 AI Ollama 入門隱私成本 <![CDATA[[AI 怎麼問 #7] AI 做不好哪些事？2026 年用 ChatGPT / Claude 前必須知道的四個地雷]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-what-ai-does-poorly https://ai-muninn.com/zh-TW/blog/ai-ask-right-what-ai-does-poorly Thu, 16 Apr 2026 00:00:00 GMT AI 幻覺 ChatGPT Claude 隱私入門 <![CDATA[[AI Agent] Gemma 4 26B 跑通 SWE-bench Lite 單題：兩天 28 次 run，2 次真的算數]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-swe-bench-scaffold-engineering https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-swe-bench-scaffold-engineering Wed, 15 Apr 2026 00:00:00 GMT SWE-bench mini-swe-agent Gemma 4 vLLM agent scaffold GX10 Gemma 4 26B-A4B <![CDATA[[LLM 深水區] 量化演算法在做什麼？從 Q4_K_M 到 TurboQuant 的三層拆解]]> https://ai-muninn.com/zh-TW/blog/llm-deep-dive-quantization-algorithms https://ai-muninn.com/zh-TW/blog/llm-deep-dive-quantization-algorithms Wed, 15 Apr 2026 00:00:00 GMT LLM 量化 Quantization K-quant TurboQuant PolarQuant QJL 演算法 <![CDATA[[AI 怎麼問 #6] 追問的藝術 — 第一個答案太淺怎麼辦]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-follow-up-questions https://ai-muninn.com/zh-TW/blog/ai-ask-right-follow-up-questions Tue, 14 Apr 2026 00:00:00 GMT AI 對話技巧追問入門工作效率 <![CDATA[[LLM 101 #5] Context Window — AI 一次能讀多少字？]]> https://ai-muninn.com/zh-TW/blog/llm-101-context-window https://ai-muninn.com/zh-TW/blog/llm-101-context-window Tue, 14 Apr 2026 00:00:00 GMT LLM Context Window 入門科普 Token AI 記憶 <![CDATA[[AI Agent] Gemma 4 從 40 次失敗到 9 步修好 Bug — 只換了一個東西]]> https://ai-muninn.com/zh-TW/blog/swe-bench-local-models-framework-matters https://ai-muninn.com/zh-TW/blog/swe-bench-local-models-framework-matters Mon, 13 Apr 2026 00:00:00 GMT SWE-Bench Gemma 4 Qwen 3.5 OpenHands SWE-agent vLLM DGX Spark Tool Calling AI Agent Gemma 4 26B-A4B Qwen 3.5 35B <![CDATA[[Benchmark] Gemma 4 全家桶 on DGX Spark — 哪個版本適合你？]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-complete-guide https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-complete-guide Mon, 13 Apr 2026 00:00:00 GMT Gemma 4 DGX Spark GB10 benchmark MoE NVFP4 vLLM Ollama DGX Spark 值得買嗎 DGX Spark vs RTX 5090 DGX Spark 最佳模型 2026 <![CDATA[Claude Code Token 燒太快？8 招讓你的 Session 撐 10 倍久]]> https://ai-muninn.com/zh-TW/blog/dev-workflow-token-burn-rate https://ai-muninn.com/zh-TW/blog/dev-workflow-token-burn-rate Mon, 13 Apr 2026 00:00:00 GMT Claude Code tokens context window 新手 tips MCP CLAUDE.md QMD knowledge graph musubi AI 省 token save tokens <![CDATA[[AI 怎麼問 #5] 在你動手做之前，先問：這東西已經有人做了嗎？]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-does-it-exist https://ai-muninn.com/zh-TW/blog/ai-ask-right-does-it-exist Mon, 13 Apr 2026 00:00:00 GMT AI 工具推薦工作效率入門研究 <![CDATA[[Claude Code] 讓 Claude 自己幫你瘦身 — 打造 /slim 設定檔健檢 Skill]]> https://ai-muninn.com/zh-TW/blog/claude-code-slim-self-audit-skill https://ai-muninn.com/zh-TW/blog/claude-code-slim-self-audit-skill Mon, 13 Apr 2026 00:00:00 GMT Claude Code tokens context window skills CLAUDE.md MEMORY.md 設定優化自我診斷 QMD AI Workflow <![CDATA[[DGX Spark] 從開箱到跑起來：完整部署指南]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-deployment-guide https://ai-muninn.com/zh-TW/blog/dgx-spark-deployment-guide Mon, 13 Apr 2026 00:00:00 GMT DGX Spark GB10 GX10 vLLM Ollama deployment Blackwell SM121 DGX Spark 值得買嗎 DGX Spark 設定 2026 DGX Spark vs RTX 5090 Gemma 4 Gemma 4 26B-A4B <![CDATA[[AI 怎麼問 #4] 為什麼你覺得 AI 沒用？答案機器 vs 協作工具]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-why-ai-feels-useless https://ai-muninn.com/zh-TW/blog/ai-ask-right-why-ai-feels-useless Sat, 11 Apr 2026 00:00:00 GMT AI ChatGPT 提問技巧入門工作流程心態 <![CDATA[[LLM 101 #4] 什麼是量化？Q4、Q8、FP16 到底差在哪]]> https://ai-muninn.com/zh-TW/blog/llm-101-what-is-quantization https://ai-muninn.com/zh-TW/blog/llm-101-what-is-quantization Fri, 10 Apr 2026 00:00:00 GMT LLM 量化 Ollama 入門科普 Q4_K_M GGUF <![CDATA[[AI 怎麼問 #3] 你不知道自己需要什麼 — 讓 AI 幫你挖出來]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-find-your-needs https://ai-muninn.com/zh-TW/blog/ai-ask-right-find-your-needs Fri, 10 Apr 2026 00:00:00 GMT AI 工作效率入門工作流程 <![CDATA[[LLM 101 #3] 那麼多模型，到底該下載哪一個？]]> https://ai-muninn.com/zh-TW/blog/llm-101-how-to-choose-a-model https://ai-muninn.com/zh-TW/blog/llm-101-how-to-choose-a-model Fri, 10 Apr 2026 00:00:00 GMT LLM 模型選擇 Ollama 入門科普 Gemma Llama Qwen <![CDATA[[AI 怎麼問 #2] 打開 AI 之後，你的第一句話該怎麼說？]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-first-message https://ai-muninn.com/zh-TW/blog/ai-ask-right-first-message Thu, 09 Apr 2026 00:00:00 GMT AI ChatGPT 入門對話技巧 <![CDATA[[AI 怎麼問 #1] 2026 年，你可以用的 AI 有哪些？]]> https://ai-muninn.com/zh-TW/blog/ai-ask-right-which-ai-to-use-2026 https://ai-muninn.com/zh-TW/blog/ai-ask-right-which-ai-to-use-2026 Thu, 09 Apr 2026 00:00:00 GMT AI ChatGPT Claude Gemini 入門 <![CDATA[[Benchmark] 拯救 Gemma 4 31B：在 32GB MacBook Pro 上從 1.5 加速到 12.8 tok/s]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-31b-rescue-mbp-32gb https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-31b-rescue-mbp-32gb Wed, 08 Apr 2026 00:00:00 GMT Gemma 4 31B M1 Max Ollama oMLX swap KV cache TurboQuant Apple Silicon memory management Gemma 4 31B <![CDATA[[Benchmark] 4 台機器、4 個模型、1 個答案：記憶體決定一切]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-4-machines-4-models-bandwidth https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-4-machines-4-models-bandwidth Wed, 08 Apr 2026 00:00:00 GMT 頻寬速度。]]> Gemma 4 RTX 5090 DGX Spark GB10 M1 Max M4 Ollama benchmark memory bandwidth swap E2B E4B 26B 31B <![CDATA[[LLM 101 #2] Dense、MoE、PLE、SSM — 四種 AI 模型架構，一次搞懂]]> https://ai-muninn.com/zh-TW/blog/llm-101-dense-moe-ple-ssm-architectures https://ai-muninn.com/zh-TW/blog/llm-101-dense-moe-ple-ssm-architectures Wed, 08 Apr 2026 00:00:00 GMT Dense MoE PLE SSM Mamba LLM 模型架構入門科普 <![CDATA[[Benchmark] Gemma 4 E2B vs E4B：三台機器實測，記憶體頻寬決定一切]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-e2b-vs-e4b-ollama-3-machines https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-e2b-vs-e4b-ollama-3-machines Tue, 07 Apr 2026 00:00:00 GMT Gemma 4 E2B E4B Ollama benchmark DGX Spark GB10 M1 Max M4 Apple Silicon 記憶體頻寬 Gemma 4 E2B Gemma 4 E4B <![CDATA[[Benchmark] 從 19 到 50 tok/s：我們搶先做了全球首個 Gemma 4 E4B NVFP4 量化]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-e4b-nvfp4-50-toks https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-e4b-nvfp4-50-toks Tue, 07 Apr 2026 00:00:00 GMT Gemma 4 E4B NVFP4 FP8 vLLM DGX Spark GB10 quantization llm-compressor PLE HuggingFace Gemma 4 E4B <![CDATA[[LLM 101 #1] Ollama vs vLLM：在自己電腦跑 AI 的兩條路]]> https://ai-muninn.com/zh-TW/blog/llm-101-ollama-vs-vllm https://ai-muninn.com/zh-TW/blog/llm-101-ollama-vs-vllm Tue, 07 Apr 2026 00:00:00 GMT Ollama vLLM LLM 本地部署入門科普 <![CDATA[[Benchmark] Gemma 4 31B Dense 跑在 DGX Spark：7 tok/s 和頻寬之牆]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-31b-dense-7-toks https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-31b-dense-7-toks Sun, 05 Apr 2026 00:00:00 GMT Gemma 4 NVFP4 vLLM DGX Spark GB10 SM121 dense benchmark bandwidth Gemma 4 31B Gemma 4 31B-IT <![CDATA[[Benchmark] 同模型 vLLM vs Ollama：為什麼 GB10 上差 30%]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-vllm-vs-ollama-same-model https://ai-muninn.com/zh-TW/blog/dgx-spark-vllm-vs-ollama-same-model Sun, 05 Apr 2026 00:00:00 GMT vLLM Ollama benchmark DGX Spark GB10 Gemma 4 NVFP4 inference Gemma 4 26B-A4B <![CDATA[[vLLM] Gemma 4 26B-A4B NVFP4 跑在 DGX Spark：52 tok/s，模型只佔 16 GB]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-26b-nvfp4-52-toks https://ai-muninn.com/zh-TW/blog/dgx-spark-gemma4-26b-nvfp4-52-toks Sun, 05 Apr 2026 00:00:00 GMT Gemma 4 NVFP4 vLLM DGX Spark GB10 SM121 MoE benchmark Gemma 4 26B-A4B <![CDATA[[DGX Spark] 過熱、100W 功耗上限、30W 安全模式 — 完整診斷指南]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-30w-power-safety-mode https://ai-muninn.com/zh-TW/blog/dgx-spark-30w-power-safety-mode Thu, 02 Apr 2026 00:00:00 GMT GX10 GB10 DGX Spark power delivery vLLM hardware overheating 100W Carmack Gemma 4 Gemma 4 26B-A4B <![CDATA[[Benchmark] TurboQuant 實測：KV Cache 3-bit 壓縮，真的零損失？]]> https://ai-muninn.com/zh-TW/blog/turboquant-kv-cache-benchmark-gx10 https://ai-muninn.com/zh-TW/blog/turboquant-kv-cache-benchmark-gx10 Mon, 30 Mar 2026 00:00:00 GMT TurboQuant KV Cache Quantization vLLM Benchmark Qwen3.5 GX10 SM121 <![CDATA[[AI Agent] openclaw + ChatGPT OAuth：不買 API 額度也能用 GPT-5.4]]> https://ai-muninn.com/zh-TW/blog/openclaw-chatgpt-oauth-gpt54-no-api-key https://ai-muninn.com/zh-TW/blog/openclaw-chatgpt-oauth-gpt54-no-api-key Tue, 24 Mar 2026 00:00:00 GMT openclaw GPT-5.4 ChatGPT OAuth AI agent <![CDATA[[AI Agent] NemoClaw 不靠雲端：把 Nemotron 換成本地 Ollama 模型]]> https://ai-muninn.com/zh-TW/blog/nemoclaw-local-inference-ollama https://ai-muninn.com/zh-TW/blog/nemoclaw-local-inference-ollama Tue, 24 Mar 2026 00:00:00 GMT NemoClaw OpenClaw OpenShell Ollama vLLM AI Agent NVIDIA GX10 Local Inference <![CDATA[NemoClaw 安裝踩坑：官方文件沒寫的 4 個必修修正]]> https://ai-muninn.com/zh-TW/blog/nemoclaw-install-gx10-from-scratch https://ai-muninn.com/zh-TW/blog/nemoclaw-install-gx10-from-scratch Mon, 23 Mar 2026 00:00:00 GMT NemoClaw OpenClaw OpenShell AI Agent NVIDIA DGX Spark GX10 GB10 <![CDATA[NemoClaw 是什麼？NVIDIA 一鍵 AI Agent 框架完整解析]]> https://ai-muninn.com/zh-TW/blog/nemoclaw-what-it-is-why-it-exists https://ai-muninn.com/zh-TW/blog/nemoclaw-what-it-is-why-it-exists Mon, 23 Mar 2026 00:00:00 GMT NemoClaw OpenClaw OpenShell AI Agent NVIDIA DGX Spark GX10 <![CDATA[[AI Agent] openclaw 用 Telegram Bot API 9.5 sendMessageDraft 做即時串流]]> https://ai-muninn.com/zh-TW/blog/openclaw-telegram-sendmessagedraft-streaming https://ai-muninn.com/zh-TW/blog/openclaw-telegram-sendmessagedraft-streaming Sat, 21 Mar 2026 00:00:00 GMT openclaw Telegram streaming Bot API GLM <![CDATA[[AI Agent] openclaw 接上 131K Context：max_tokens 變負數的那一刻]]> https://ai-muninn.com/zh-TW/blog/openclaw-context-budget-negative-maxtokens https://ai-muninn.com/zh-TW/blog/openclaw-context-budget-negative-maxtokens Sat, 21 Mar 2026 00:00:00 GMT openclaw context window vLLM gpt-oss configuration <![CDATA[[vLLM] GB10 上的 FP8 KV Cache：為什麼輸出會在 500 Token 後崩成重複迴圈]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-fp8-kvcache-repetition https://ai-muninn.com/zh-TW/blog/dgx-spark-fp8-kvcache-repetition Sat, 21 Mar 2026 00:00:00 GMT vLLM FP8 KV cache GB10 DGX Spark quantization SM121 Qwen 3.5 Qwen 3.5 35B <![CDATA[[Claude Code] claude-agent-sdk vs subprocess：中間 Turn 為什麼消失了]]> https://ai-muninn.com/zh-TW/blog/claude-code-agent-sdk-orchestrator https://ai-muninn.com/zh-TW/blog/claude-code-agent-sdk-orchestrator Sat, 21 Mar 2026 00:00:00 GMT Claude Code claude-agent-sdk multi-agent orchestrator Python asyncio <![CDATA[[AI Agent] openclaw：Bot 突然消失了 — Tailscale、IPv6、和一個 Node.js 的安靜陷阱]]> https://ai-muninn.com/zh-TW/blog/openclaw-telegram-ipv6-tailscale-silent-bot https://ai-muninn.com/zh-TW/blog/openclaw-telegram-ipv6-tailscale-silent-bot Thu, 19 Mar 2026 00:00:00 GMT Node.js Tailscale IPv6 undici Happy Eyeballs Telegram Debugging Networking <![CDATA[[vLLM] 在 DGX Spark 上跑 120B 模型到 60 tok/s——零 API 成本、六個坑]]> https://ai-muninn.com/zh-TW/blog/part2-gpt-oss-120b-serve-script https://ai-muninn.com/zh-TW/blog/part2-gpt-oss-120b-serve-script Thu, 19 Mar 2026 00:00:00 GMT DGX Spark SM121 vLLM gpt-oss MXFP4 Blackwell LLM Serving <![CDATA[[vLLM] Qwen3.5-122B 跑起來了。但只有 14 tok/s。]]> https://ai-muninn.com/zh-TW/blog/part2-qwen-122b-14-toks-gdn-kernel-gap https://ai-muninn.com/zh-TW/blog/part2-qwen-122b-14-toks-gdn-kernel-gap Thu, 19 Mar 2026 00:00:00 GMT DGX Spark SM121 Qwen3.5-122B vLLM NVFP4 Marlin GDN LLM Serving <![CDATA[[AI Agent] openclaw：Agent 卡住的時候，叫 CLI 來幫]]> https://ai-muninn.com/zh-TW/blog/openclaw-callhelp-spawning-cli-from-agent-loop https://ai-muninn.com/zh-TW/blog/openclaw-callhelp-spawning-cli-from-agent-loop Wed, 18 Mar 2026 00:00:00 GMT AI Agent openclaw Codex LLM Agent Tools Local AI <![CDATA[[vLLM] 為什麼你的 DGX Spark 只會輸出「!!!!!」：SM121 上的 NVFP4 除錯記錄]]> https://ai-muninn.com/zh-TW/blog/part1-why-your-dgx-spark-says-exclamation-marks https://ai-muninn.com/zh-TW/blog/part1-why-your-dgx-spark-says-exclamation-marks Tue, 17 Mar 2026 00:00:00 GMT DGX Spark SM121 vLLM NVFP4 MXFP4 Blackwell CUDA LLM Serving <![CDATA[[AI Agent] Codex-Executor 模式：讓 Agent Session 保持輕量]]> https://ai-muninn.com/zh-TW/blog/openclaw-codex-executor-agent-architecture https://ai-muninn.com/zh-TW/blog/openclaw-codex-executor-agent-architecture Mon, 16 Mar 2026 00:00:00 GMT AI Agent Claude Code Codex Agent Architecture OpenClaw <![CDATA[[vLLM] 單顆 GB10 跑 Nemotron-3-Super-120B：一天的除錯記錄]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-nemotron-120b-vllm https://ai-muninn.com/zh-TW/blog/dgx-spark-nemotron-120b-vllm Fri, 13 Mar 2026 00:00:00 GMT DGX Spark GB10 SM121 Nemotron vLLM NVFP4 Blackwell LLM Serving Nemotron 3 Super Nemotron 3 Super 120B <![CDATA[[vLLM] Ollama 的 KEEP_ALIVE 在偷吃你的 vLLM 記憶體空間]]> https://ai-muninn.com/zh-TW/blog/openclaw-ollama-vllm-gpu-conflict https://ai-muninn.com/zh-TW/blog/openclaw-ollama-vllm-gpu-conflict Sat, 07 Mar 2026 00:00:00 GMT vLLM Ollama GPU Memory DGX Spark GB10 LLM Serving <![CDATA[[vLLM] SSM 模型不能加 --enable-chunked-prefill]]> https://ai-muninn.com/zh-TW/blog/openclaw-chunked-prefill-ssm-trap https://ai-muninn.com/zh-TW/blog/openclaw-chunked-prefill-ssm-trap Fri, 06 Mar 2026 00:00:00 GMT vLLM SSM Qwen DGX Spark LLM Serving Performance <![CDATA[[vLLM] Qwen3.5-35B 跑到 47 tok/s：從 Ollama 遷移到 vLLM]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-vllm-qwen35-setup https://ai-muninn.com/zh-TW/blog/dgx-spark-vllm-qwen35-setup Thu, 05 Mar 2026 00:00:00 GMT DGX Spark GB10 vLLM Ollama Qwen 3.5 Docker Blackwell AI Agent Qwen 3.5 35B <![CDATA[[AI Agent] 零 API 成本：用 DGX Spark + Mac Mini 跑 OpenClaw]]> https://ai-muninn.com/zh-TW/blog/openclaw-dgx-spark-local-ai-agent https://ai-muninn.com/zh-TW/blog/openclaw-dgx-spark-local-ai-agent Thu, 05 Mar 2026 00:00:00 GMT OpenClaw AI Agent DGX Spark Mac Mini Self-Hosted Ollama SearXNG <![CDATA[[Benchmark] 純 MoE vs SSM Hybrid：Context Decay 與為什麼 Agent 要在乎]]> https://ai-muninn.com/zh-TW/blog/openclaw-moe-ssm-context-decay https://ai-muninn.com/zh-TW/blog/openclaw-moe-ssm-context-decay Sun, 01 Mar 2026 00:00:00 GMT Benchmark SSM MoE DGX Spark GB10 LLM Serving AI Agents <![CDATA[[Dev Workflow] 讓兩個 AI 吵架。它們不同意的地方才是重點。]]> https://ai-muninn.com/zh-TW/blog/claude-code-debate-system https://ai-muninn.com/zh-TW/blog/claude-code-debate-system Thu, 26 Feb 2026 00:00:00 GMT Dev Workflow Claude Code Gemini Codex Multi-AI Code Review <![CDATA[[Claude Code] 用 Claude Code 測 iOS App：context 用量砍 81%]]> https://ai-muninn.com/zh-TW/blog/claude-code-ios-testing-bpstracker https://ai-muninn.com/zh-TW/blog/claude-code-ios-testing-bpstracker Thu, 26 Feb 2026 00:00:00 GMT Claude Code iOS Swift Testing Fastlane BPS Tracker <![CDATA[[AI Agent] OpenClaw Config 熱重載：不需要重啟]]> https://ai-muninn.com/zh-TW/blog/openclaw-config-hot-reload https://ai-muninn.com/zh-TW/blog/openclaw-config-hot-reload Wed, 25 Feb 2026 00:00:00 GMT AI Agent OpenClaw Configuration Developer Workflow <![CDATA[[Claude Code] 我寫了 MANDATORY。AI 還是沒跑。]]> https://ai-muninn.com/zh-TW/blog/claude-code-mandatory-instructions https://ai-muninn.com/zh-TW/blog/claude-code-mandatory-instructions Thu, 19 Feb 2026 00:00:00 GMT Claude Code AI Agents Prompt Engineering Systems Design Developer Workflow <![CDATA[[Benchmark] DGX Spark 跑 8 個模型：找出最適合 AI Agent 的組合]]> https://ai-muninn.com/zh-TW/blog/dgx-spark-ollama-benchmark-8-models https://ai-muninn.com/zh-TW/blog/dgx-spark-ollama-benchmark-8-models Thu, 19 Feb 2026 00:00:00 GMT DGX Spark GB10 Ollama Benchmark LLM AI Agent Blackwell