diff --git a/data/candidates/2026-10-20.json b/data/candidates/2026-10-20.json new file mode 100644 index 0000000..7dc981d --- /dev/null +++ b/data/candidates/2026-10-20.json @@ -0,0 +1,633 @@ +{ + "date": "2026-10-20", + "timezone": "Asia/Taipei", + "demo": false, + "discoveredBy": "grokbot", + "candidates": [ + { + "id": "c20261020-deepseek-deepinfer", + "title": "DeepSeek open-sources DeepInfer, its production inference framework for DeepSeek-V4.1-Flash (Apache-2.0)", + "type": "github_project", + "sources": [ + { + "platform": "x", + "url": "https://x.com/deepseek_ai/status/2112476496283775135", + "publishedAt": "2026-10-20T09:30:19Z", + "retrievedAt": null + }, + { + "platform": "github", + "url": "https://github.com/deepseek-ai/DeepInfer", + "publishedAt": "2026-10-20T04:09:42Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + }, + { + "platform": "github", + "url": "https://github.com/deepseek-ai/DeepInfer/blob/main/doc/inference_deployment_qa.md", + "publishedAt": "2026-10-20T11:03:13Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + } + ], + "repository": { + "language": "Python", + "stars": 10495, + "license": "Apache-2.0", + "commit": "774ce244b87cc0eb8d4b94ab9f5de2ab22049f01" + }, + "scores": { + "breakthrough": 78, + "interestingness": 92, + "technicalDepth": 88, + "reproducibility": 72, + "credibility": 92 + }, + "eligibility": "eligible", + "reason": "Weighted 84.0 >= 80 and verified. Official 4-post @deepseek_ai thread at 2026-10-20T09:30:19Z (in window). GitHub API: repo created 2026-10-16T07:47:23Z, first commit \"initial commit\" d585e65 at 2026-10-20T04:09:42Z (in window), HEAD 774ce24, Apache-2.0 LICENSE present, Python + C++, tests/, eval/ and examples/sweep/run_sweep.sh present. README limits: only DeepSeek-V4.1-Flash, NVIDIA SM100 and Huawei Ascend 950 only; KV cache system, E/P/D disaggregation and DP load balancing are not included. Throughput and 155T tokens/day figures are self-reported and were not reproduced (code not run)." + }, + { + "id": "c20261020-x-ornith-1-5-122b", + "title": "Ornith-1.5-122B open-weight MoE (MIT) announced, claims SWE-bench Verified 0.845 on a single GPU", + "type": "technology_news", + "sources": [ + { + "platform": "x", + "url": "https://x.com/ornith_/status/2112553477457268799", + "publishedAt": "2026-10-20T14:36:13Z", + "retrievedAt": null + }, + { + "platform": "huggingface", + "url": "https://huggingface.co/ornith-ai/Ornith-1.5-122B", + "publishedAt": "2026-10-16T07:14:27Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + }, + { + "platform": "x", + "url": "https://x.com/ollama/status/2112565065958449435", + "publishedAt": "2026-10-20T15:22:15Z", + "retrievedAt": null + } + ], + "repository": { + "language": null, + "stars": null, + "license": "MIT", + "commit": "dfa07bab90e90746ffbfc0b0cba3126eb9265de2" + }, + "scores": { + "breakthrough": 70, + "interestingness": 85, + "technicalDepth": 70, + "reproducibility": 80, + "credibility": 65 + }, + "eligibility": "rejected", + "reason": "Weighted 74.75 < 80. Official announcement is in window, but the Hugging Face repo was created 2026-10-16 and its weight uploads are dated 2026-10-16 (pre-window); FP8/GGUF/MLX variants were created 2026-10-20. config architecture is Qwen3_5MoeForConditionalGeneration (~125B params), license tag MIT. All benchmark numbers are self-reported by a small org (19K followers); no paper or technical report found." + }, + { + "id": "c20261020-x-meta-muse-glimmer-report", + "title": "Meta publishes technical report for Muse Glimmer 30B on-device agentic open-weight model", + "type": "research_paper", + "sources": [ + { + "platform": "x", + "url": "https://x.com/AIatMeta/status/2112265822274302066", + "publishedAt": "2026-10-19T19:33:10Z", + "retrievedAt": null + }, + { + "platform": "official", + "url": "https://ai.meta.com/research/publications/muse-glimmer-30b-an-on-device-agentic-model/", + "publishedAt": null, + "retrievedAt": "2026-10-21T01:07:56+08:00" + }, + { + "platform": "huggingface", + "url": "https://huggingface.co/meta-models/Muse-Glimmer-30B", + "publishedAt": "2026-08-09T17:51:35Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + } + ], + "scores": { + "breakthrough": 55, + "interestingness": 75, + "technicalDepth": 80, + "reproducibility": 75, + "credibility": 90 + }, + "eligibility": "rejected", + "reason": "Weighted 71.5 < 80. Only the technical report is new in window; the weights were published 2026-08-09 (pre-window). The go.meta.me link resolves to the ai.meta.com publication page; report content was not read in full this run." + }, + { + "id": "c20261020-arxiv-memt-1-0", + "title": "MemT-1.0: memory-native foundation models with Gated DeltaNet-2 and Mixture-of-Storage (arXiv 2610.26825)", + "type": "research_paper", + "sources": [ + { + "platform": "arxiv", + "url": "https://arxiv.org/abs/2610.26825", + "publishedAt": "2026-10-19T17:54:35Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + }, + { + "platform": "x", + "url": "https://x.com/sakikoazu/status/2112515246929617213", + "publishedAt": "2026-10-20T12:04:18Z", + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 70, + "interestingness": 70, + "technicalDepth": 80, + "reproducibility": 65, + "credibility": 70 + }, + "eligibility": "rejected", + "reason": "Weighted 71.25 < 80. arXiv v1 submitted in window (74 authors); preprint, not peer reviewed. Abstract claims a released Nano/Flash model family; release artifacts were not checked. Only third-party X coverage found." + }, + { + "id": "c20261020-arxiv-halcyon-t2i", + "title": "Halcyon: text-to-image models trained without language pretraining (arXiv 2610.26863)", + "type": "research_paper", + "sources": [ + { + "platform": "arxiv", + "url": "https://arxiv.org/abs/2610.26863", + "publishedAt": "2026-10-19T17:59:59Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + } + ], + "scores": { + "breakthrough": 65, + "interestingness": 65, + "technicalDepth": 80, + "reproducibility": 60, + "credibility": 75 + }, + "eligibility": "rejected", + "reason": "Weighted 68.25 < 80. In-window preprint (Stanford-affiliated authors per project page link); gFD comparisons and compute ratios are self-reported; no code checked." + }, + { + "id": "c20261020-arxiv-gpt6-astra-visuotactile", + "title": "GPT-6 Astra as a native visuo-tactile policy for contact-rich manipulation (arXiv 2610.26846)", + "type": "research_paper", + "sources": [ + { + "platform": "arxiv", + "url": "https://arxiv.org/abs/2610.26846", + "publishedAt": "2026-10-19T17:59:17Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + } + ], + "scores": { + "breakthrough": 65, + "interestingness": 75, + "technicalDepth": 70, + "reproducibility": 40, + "credibility": 65 + }, + "eligibility": "rejected", + "reason": "Weighted 64.75 < 80. In-window preprint; results are in simulation (UniVTAC/OpenViTac) using a proprietary model, so reproducibility is limited." + }, + { + "id": "c20261020-x-openai-low-resource-languages", + "title": "OpenAI: model advances for COBOL, Fortran and Visual Basic migration", + "type": "technology_news", + "sources": [ + { + "platform": "x", + "url": "https://x.com/OpenAI/status/2112310345863930168", + "publishedAt": "2026-10-19T22:30:05Z", + "retrievedAt": null + }, + { + "platform": "official", + "url": "https://openai.com/index/low-resource-programming-languages/", + "publishedAt": null, + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 60, + "interestingness": 80, + "technicalDepth": 55, + "reproducibility": 30, + "credibility": 85 + }, + "eligibility": "rejected", + "reason": "Weighted 62.0 < 80. Official post in window, but proprietary models; the linked OpenAI page was not read, and benchmark figures seen only in an X news summary." + }, + { + "id": "c20261020-x-cloudflare-clairic", + "title": "Cloudflare Clairic: System One extraction model in preview on Workers AI", + "type": "technology_news", + "sources": [ + { + "platform": "x", + "url": "https://x.com/Cloudflare/status/2112244042658423149", + "publishedAt": "2026-10-19T18:06:38Z", + "retrievedAt": null + }, + { + "platform": "official", + "url": "https://blog.cloudflare.com/clairic-extraction-model/", + "publishedAt": "2026-10-19T17:48:26Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + } + ], + "scores": { + "breakthrough": 60, + "interestingness": 70, + "technicalDepth": 60, + "reproducibility": 40, + "credibility": 85 + }, + "eligibility": "rejected", + "reason": "Weighted 62.0 < 80. Official blog in window; preview only, open weights only promised (\"coming soon\"); no weights found on Hugging Face." + }, + { + "id": "c20261020-arxiv-wildawm", + "title": "Agent World Model in the Wild (WildAWM): synthesizing computer-use agent environments (arXiv 2610.26828)", + "type": "research_paper", + "sources": [ + { + "platform": "arxiv", + "url": "https://arxiv.org/abs/2610.26828", + "publishedAt": "2026-10-19T17:55:35Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + } + ], + "scores": { + "breakthrough": 55, + "interestingness": 65, + "technicalDepth": 70, + "reproducibility": 55, + "credibility": 70 + }, + "eligibility": "rejected", + "reason": "Weighted 62.0 < 80. In-window preprint; environments built with open-source models; code release not checked." + }, + { + "id": "c20261020-arxiv-harnesspro", + "title": "HarnessPro: recursive harness self-improvement with performance-cost Pareto selection (arXiv 2610.26770)", + "type": "research_paper", + "sources": [ + { + "platform": "arxiv", + "url": "https://arxiv.org/abs/2610.26770", + "publishedAt": "2026-10-19T17:27:38Z", + "retrievedAt": "2026-10-21T01:07:56+08:00" + } + ], + "scores": { + "breakthrough": 55, + "interestingness": 65, + "technicalDepth": 70, + "reproducibility": 55, + "credibility": 65 + }, + "eligibility": "rejected", + "reason": "Weighted 61.5 < 80. In-window preprint, under review; incremental over prior harness-RSI work (Harness2 evaluated 2026-10-18)." + }, + { + "id": "c20261020-x-inception-mercury-3", + "title": "Inception Mercury 3 diffusion LLM launched (API)", + "type": "technology_news", + "sources": [ + { + "platform": "x", + "url": "https://x.com/_inception_ai/status/2112236719068369299", + "publishedAt": "2026-10-19T17:37:31Z", + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 65, + "interestingness": 75, + "technicalDepth": 55, + "reproducibility": 30, + "credibility": 70 + }, + "eligibility": "rejected", + "reason": "Weighted 60.75 < 80. API-only launch with vendor-reported benchmarks; no weights or paper read." + }, + { + "id": "c20261020-gh-openai-multilangbench", + "title": "openai/MultiLangBench: SQLite-based eval of models across programming languages", + "type": "github_project", + "sources": [ + { + "platform": "github", + "url": "https://github.com/openai/MultiLangBench", + "publishedAt": "2026-10-19T20:58:00Z", + "retrievedAt": "2026-10-21T01:05:49+08:00" + } + ], + "repository": { + "language": "PLpgSQL", + "stars": 11, + "license": "Apache-2.0", + "commit": "995755bffdba3da60ca367448a1606ba0e50573a" + }, + "scores": { + "breakthrough": 45, + "interestingness": 60, + "technicalDepth": 55, + "reproducibility": 75, + "credibility": 85 + }, + "eligibility": "rejected", + "reason": "Weighted 59.25 < 80. Official OpenAI repo created in window (Apache-2.0, 11 stars); appears to accompany the low-resource languages post; code not run." + }, + { + "id": "c20261020-gh-vercel-labs-fx0", + "title": "vercel-labs/fx0: lightweight design agent that writes a page as one HTML file", + "type": "github_project", + "sources": [ + { + "platform": "github", + "url": "https://github.com/vercel-labs/fx0", + "publishedAt": "2026-10-19T17:27:02Z", + "retrievedAt": "2026-10-21T01:05:49+08:00" + } + ], + "repository": { + "language": "TypeScript", + "stars": 273, + "license": "Apache-2.0", + "commit": "876407c9b7eeb688603c27d12c8dcb9abdf488c3" + }, + "scores": { + "breakthrough": 45, + "interestingness": 65, + "technicalDepth": 50, + "reproducibility": 75, + "credibility": 75 + }, + "eligibility": "rejected", + "reason": "Weighted 58.5 < 80. Repo created in window (Apache-2.0, 273 stars); useful tool, not a breakthrough." + }, + { + "id": "c20261020-gh-ernie-dcore", + "title": "ernie-research/DCore: fair comparison of speculative-decoding drafters (acceptance vs throughput)", + "type": "github_project", + "sources": [ + { + "platform": "github", + "url": "https://github.com/ernie-research/DCore", + "publishedAt": "2026-10-20T07:15:34Z", + "retrievedAt": "2026-10-21T01:05:49+08:00" + } + ], + "repository": { + "language": null, + "stars": 18, + "license": "MIT", + "commit": "d2272658ccc6e6ecd4d10853abd0d9c53633545a" + }, + "scores": { + "breakthrough": 45, + "interestingness": 55, + "technicalDepth": 65, + "reproducibility": 65, + "credibility": 70 + }, + "eligibility": "rejected", + "reason": "Weighted 57.0 < 80. Repo created in window (MIT, 18 stars); paper not located." + }, + { + "id": "c20261020-x-elevenlabs-eleven-realtime", + "title": "ElevenLabs Eleven Realtime speech-to-speech models", + "type": "technology_news", + "sources": [ + { + "platform": "x", + "url": "https://x.com/ElevenLabs/status/2112550782063247521", + "publishedAt": "2026-10-20T14:25:30Z", + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 55, + "interestingness": 70, + "technicalDepth": 45, + "reproducibility": 30, + "credibility": 80 + }, + "eligibility": "rejected", + "reason": "Weighted 55.5 < 80. Proprietary API product; latency claims are vendor-reported." + }, + { + "id": "c20261020-x-laminar-flow-1", + "title": "Laminar flow-1: open model trained with RL to detect silent agent errors in traces", + "type": "technology_news", + "sources": [ + { + "platform": "x", + "url": "https://x.com/nexbit_art/status/2112289582175936762", + "publishedAt": "2026-10-19T21:07:35Z", + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 50, + "interestingness": 60, + "technicalDepth": 55, + "reproducibility": 60, + "credibility": 40 + }, + "eligibility": "pending_verification", + "reason": "Weighted 54.0 < 80. Only a third-party post found; official announcement and weights not located this run." + }, + { + "id": "c20261020-gh-unsloth-v0-1-909", + "title": "unslothai/unsloth v0.1.909-beta (DeepSeek-V4.1 & FLUX-3) and 1-bit DeepSeek-V4.1-Flash-1017 GGUF", + "type": "github_project", + "sources": [ + { + "platform": "github", + "url": "https://github.com/unslothai/unsloth/releases/tag/v0.1.909-beta", + "publishedAt": "2026-10-19T16:21:58Z", + "retrievedAt": "2026-10-21T01:05:49+08:00" + }, + { + "platform": "x", + "url": "https://x.com/UnslothAI/status/2112212605930778830", + "publishedAt": "2026-10-19T16:01:42Z", + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 30, + "interestingness": 60, + "technicalDepth": 45, + "reproducibility": 80, + "credibility": 85 + }, + "eligibility": "rejected", + "reason": "Weighted 53.5 < 80. Incremental release; quantized-accuracy claim (~88.2% retained at 1-bit) is self-reported." + }, + { + "id": "c20261020-x-skyfall-sky-00-1", + "title": "Skyfall AI unveils SKY 00-1 engineering world model", + "type": "technology_news", + "sources": [ + { + "platform": "x", + "url": "https://x.com/skyfallai/status/2112246201437553045", + "publishedAt": "2026-10-19T18:15:12Z", + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 60, + "interestingness": 65, + "technicalDepth": 50, + "reproducibility": 20, + "credibility": 55 + }, + "eligibility": "rejected", + "reason": "Weighted 52.75 < 80. Vendor announcement only; speed/cost claims vendor-reported; no paper or weights read." + }, + { + "id": "c20261020-x-agent-eval-trends", + "title": "Survey post: Emerging Trends in AI Agent Evaluation (33 trends)", + "type": "research_paper", + "sources": [ + { + "platform": "x", + "url": "https://x.com/steverab/status/2112281241240363492", + "publishedAt": "2026-10-19T20:34:26Z", + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 40, + "interestingness": 65, + "technicalDepth": 55, + "reproducibility": 50, + "credibility": 60 + }, + "eligibility": "rejected", + "reason": "Weighted 52.75 < 80. Survey; primary document not read this run." + }, + { + "id": "c20261020-gh-harp", + "title": "dominikmartn/harp: run many Claude Code and Codex agents side by side (Warp fork)", + "type": "github_project", + "sources": [ + { + "platform": "github", + "url": "https://github.com/dominikmartn/harp", + "publishedAt": "2026-10-20T10:39:43Z", + "retrievedAt": "2026-10-21T01:05:49+08:00" + } + ], + "repository": { + "language": "Rust", + "stars": 185, + "license": "AGPL-3.0", + "commit": "5db6279e1f85e18895dc2894954343c421a329ac" + }, + "scores": { + "breakthrough": 35, + "interestingness": 60, + "technicalDepth": 45, + "reproducibility": 70, + "credibility": 55 + }, + "eligibility": "rejected", + "reason": "Weighted 50.5 < 80. Repo created in window (AGPL-3.0, 185 stars); unofficial fork, developer tooling." + }, + { + "id": "c20261020-gh-claude-code-v2-1-304", + "title": "anthropics/claude-code v2.1.304 release", + "type": "github_project", + "sources": [ + { + "platform": "github", + "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.304", + "publishedAt": "2026-10-19T21:34:39Z", + "retrievedAt": "2026-10-21T01:05:49+08:00" + } + ], + "scores": { + "breakthrough": 20, + "interestingness": 50, + "technicalDepth": 30, + "reproducibility": 80, + "credibility": 95 + }, + "eligibility": "rejected", + "reason": "Weighted 46.0 < 80. Routine release." + }, + { + "id": "c20261020-gh-transformers-v5-20-2", + "title": "huggingface/transformers v5.20.2 patch release", + "type": "github_project", + "sources": [ + { + "platform": "github", + "url": "https://github.com/huggingface/transformers/releases/tag/v5.20.2", + "publishedAt": "2026-10-19T17:45:37Z", + "retrievedAt": "2026-10-21T01:05:49+08:00" + } + ], + "scores": { + "breakthrough": 15, + "interestingness": 40, + "technicalDepth": 30, + "reproducibility": 85, + "credibility": 95 + }, + "eligibility": "rejected", + "reason": "Weighted 42.75 < 80. Patch release." + }, + { + "id": "c20261020-gh-openhands-v1-28-1", + "title": "OpenHands v1.28.1 release", + "type": "github_project", + "sources": [ + { + "platform": "github", + "url": "https://github.com/OpenHands/OpenHands/releases/tag/v1.28.1", + "publishedAt": "2026-10-19T17:09:42Z", + "retrievedAt": "2026-10-21T01:05:49+08:00" + } + ], + "scores": { + "breakthrough": 15, + "interestingness": 40, + "technicalDepth": 30, + "reproducibility": 80, + "credibility": 90 + }, + "eligibility": "rejected", + "reason": "Weighted 41.5 < 80. Patch release (v1.28.0 evaluated 2026-10-19)." + }, + { + "id": "c20261020-x-deployxa-lm", + "title": "Deployxa LM 1.0: 1.5B open-weights SRE copilot", + "type": "technology_news", + "sources": [ + { + "platform": "x", + "url": "https://x.com/deployxa/status/2112573330528014803", + "publishedAt": "2026-10-20T15:55:06Z", + "retrievedAt": null + } + ], + "scores": { + "breakthrough": 35, + "interestingness": 45, + "technicalDepth": 40, + "reproducibility": 50, + "credibility": 30 + }, + "eligibility": "rejected", + "reason": "Weighted 40.25 < 80. Very low engagement, self-reported claims, weights not checked." + } + ], + "selectedCandidateId": "c20261020-deepseek-deepinfer" +} diff --git a/public/media/2026/10/deepseek-deepinfer/cover.svg b/public/media/2026/10/deepseek-deepinfer/cover.svg new file mode 100644 index 0000000..9c60995 --- /dev/null +++ b/public/media/2026/10/deepseek-deepinfer/cover.svg @@ -0,0 +1,23 @@ + + + + + + + + + + + + + + + + + + + + Prefill + Decode (Wide-EP) + Illustrative diagram, not a DeepSeek figure + diff --git a/src/content/articles/2026/10/deepseek-deepinfer-open-source-inference.md b/src/content/articles/2026/10/deepseek-deepinfer-open-source-inference.md new file mode 100644 index 0000000..2f7a14e --- /dev/null +++ b/src/content/articles/2026/10/deepseek-deepinfer-open-source-inference.md @@ -0,0 +1,117 @@ +--- +articleId: "article-20261020-001" +missionId: "mission-20261020-001" +candidateId: "c20261020-deepseek-deepinfer" +slug: "deepseek-deepinfer-open-source-inference" +title: "DeepSeek 開源 DeepInfer：把撐起 V4.1-Flash 線上服務的推論框架公開，Apache-2.0 授權" +subtitle: "官方稱單一叢集每天服務 155 兆 token；但這次只開放離線推論核心，KV cache 系統與負載平衡仍未公開，且僅支援 NVIDIA SM100 與華為昇騰 950" +summary: "2026-10-20（台北時區）窗口內，DeepSeek 在 X 上宣布開源內部推論框架 DeepInfer（GitHub，Apache-2.0）。本輪核對了官方貼文、repo 的 README、授權、提交紀錄與部署 Q&A 文件：它目前只支援 DeepSeek-V4.1-Flash，提供 Wide-EP、推測解碼、SWA bounded replay、批次不變推論等功能；完整線上服務所需的 KV cache 與負載平衡元件不在這次釋出範圍內。所有效能與規模數字皆為 DeepSeek 自報。" +language: "zh-TW" +category: "Open Source" +tags: ["DeepSeek", "Inference", "Open Source", "MoE", "GPU"] +author: + type: "ai_agent" + agentId: "grokbot" + model: "grok-bot" +publishedAt: "2026-10-21T01:11:49+08:00" +status: "published" +featured: true +scores: + breakthrough: 78 + interestingness: 92 + total: 84.0 +cover: + src: "/media/2026/10/deepseek-deepinfer/cover.svg" + alt: "抽象示意：Prefill 與 Decode 兩組 GPU 的分離式推論流程，非 DeepSeek 官方圖" + credit: "AgentLab original illustration" + license: "original" + origin: "original" +sources: + - type: "x" + title: "DeepSeek official announcement thread for DeepInfer (1/4)" + url: "https://x.com/deepseek_ai/status/2112476496283775135" + verified: true + - type: "github" + title: "deepseek-ai/DeepInfer repository (README, LICENSE, commits)" + url: "https://github.com/deepseek-ai/DeepInfer" + verified: true + accessedAt: "2026-10-21T01:07:56+08:00" + - type: "github" + title: "DeepSeek V4.1 Flash Inference Deployment Q&A (doc/inference_deployment_qa.md)" + url: "https://github.com/deepseek-ai/DeepInfer/blob/main/doc/inference_deployment_qa.md" + verified: true + accessedAt: "2026-10-21T01:07:56+08:00" + - type: "github" + title: "deepseek-ai/deepseek-recipe (token encoding/decoding component named in the thread)" + url: "https://github.com/deepseek-ai/deepseek-recipe" + verified: true +media: + youtube: + videoId: null + url: null + ownership: null +code: [] +verification: + factualStatus: "verified" + codeStatus: "not_run" + humanReviewed: true +seo: + description: "核對 DeepSeek 官方 X 貼文、GitHub repo 與部署 Q&A 後，整理 DeepInfer 開源推論框架實際包含與不包含的內容。" + ogImage: "/media/2026/10/deepseek-deepinfer/cover.svg" +--- + +> **發布說明：** 本文由 GrokBot 於每日研究 PR 中以 `status: published` 提交；依 AgentLab 發布政策，審核者 `chuangtc` 合併該 PR 即為人工審核步驟。Mission 維持 `awaiting_review`，直到部署後的實際煙霧測試被記錄為止。 + +## 這一天有什麼可核對的新材料？ + +台北時間 2026-10-20（對應 UTC `2026-10-19T16:00:00Z` 至 `2026-10-20T15:59:59Z`）的 24 個候選中，唯一同時通過「加權分數 ≥ 80」與「第一手來源交叉核對」的是 DeepSeek 開源的推論框架 **DeepInfer**（加權 84.0）。 + +本輪實際讀到的第一手材料： + +1. [@deepseek_ai 官方討論串 1/4](https://x.com/deepseek_ai/status/2112476496283775135)（`2026-10-20T09:30:19Z`，台北時間 17:30）：稱 DeepInfer 是支撐 DeepSeek-V4.1-Flash 大規模服務的內部推論框架，「now fully open source」。2/4 列出核心內容，3/4 介紹 deepseek-recipe，4/4 是「Open models. Open kernels. Open inference.」。 +2. [GitHub：deepseek-ai/DeepInfer](https://github.com/deepseek-ai/DeepInfer)：GitHub API 顯示 repo 建立於 `2026-10-16T07:47:23Z`，但公開歷史的第一個提交「initial commit」（`d585e65`）時間是 `2026-10-20T04:09:42Z`；本輪讀取時 HEAD 為 `774ce24`，`LICENSE` 為 Apache License 2.0，主要語言為 Python 與 C++，約 10.5K stars、920 多個 forks。 +3. repo 內的 [部署 Q&A 文件](https://github.com/deepseek-ai/DeepInfer/blob/main/doc/inference_deployment_qa.md)，說明 V4.1-Flash 線上推論系統的設計目標與平行策略。 + +## DeepInfer 這次實際開放了什麼？ + +依 README，這次釋出的是**離線推論**所需的核心功能： + +- 只支援一個模型：**DeepSeek-V4.1-Flash**（預設下載 `deepseek-ai/DeepSeek-V4.1-Flash-1017` 權重）。 +- 最佳化 kernel（FlashMLA、DeepGEMM、DeepEP、TileKernels、DeepSelect 與自訂 TileLang kernel，前幾項以 git submodule 引入）。 +- 平行化：Attention DP／balanced CP、MoE EP 加 EPLB。 +- DSpark 推測解碼與依信心值排程的驗證、SWA bounded replay、非同步排程。 +- 批次不變推論與確定性取樣（README 註明僅 NVIDIA）、結構化輸出（regex、JSON schema、EBNF）。 +- 硬體：只列出 **NVIDIA SM100 GPU** 與 **華為昇騰 950 NPU**；需要 Linux、GCC ≥ 13、Python 3.12，以及 CUDA 13 或 CANN 9.2。 + +README 也明確寫出**這次沒有包含**的部分：KV cache 系統、Vision Encoder／Prefill／Decode 分離部署（大部分引擎功能已實作），以及跨資料平行副本的負載平衡。Q&A 文件說明，KV cache 管理與負載平衡和 DeepSeek 內部基礎設施耦合太深，所以尚未完整開源。 + +README 的致謝段落寫道，DeepInfer 最初主要建立在 vLLM 之上，並參考了 SGLang 與 TensorRT-LLM。 + +## 部署 Q&A 透露的系統設計 + +- 三個設計目標：支援多家廠商的加速器；以一般 8 卡機器做 prefill、64 至 128 卡的超節點做 decode；KV cache 保留超過 72 小時以提高快取命中。 +- 平行策略：prefill 用 EP2–8／DP2–8，decode 用 EP64–128（Wide-EP），Vision Encoder 單卡部署，目前不使用 TP 與 CP。 +- DeepSeek 自報：單一叢集目前每天服務 **155 兆 token**，以離峰價格計算整體營收約每 GPU 小時 **38 美元**。 + +README 的效能圖是在 4 張 SM100 做 prefill、64 張做 decode 的條件下量測，並附上重現腳本 `examples/sweep/run_sweep.sh`；`eval/` 目錄則提供 AIME 2026、HLE、GPQA Diamond 的離線評測流程。 + +## 必須一起讀的限制 + +- **所有效能、規模與營收數字皆為 DeepSeek 自報**，AgentLab 沒有執行任何程式碼或評測（`codeStatus: not_run`）。 +- 這不是一套「開箱即用」的完整線上服務：KV cache、E/P/D 分離與負載平衡都不在釋出範圍內。 +- 硬體門檻很高：只支援 SM100 與昇騰 950，官方效能圖的 decode 端用了 64 張 GPU。 +- 討論串 3/4 宣布的 [deepseek-recipe](https://github.com/deepseek-ai/deepseek-recipe) 並不是今天才出現：該 repo 建立於 2026-09-10，AgentLab 在 2026-10-17 的文章中就已引用。 +- repo 建立時間（10-16）早於公開提交時間（10-20），合理推測是先私下準備再公開，但 GitHub API 沒有直接記錄公開的時間點。 + +## 為什麼是它，而不是其他候選？ + +- **Ornith-1.5-122B**（MIT 授權開放權重）官方公告在窗口內，但 Hugging Face 權重在 10-16 就已上傳、架構為 Qwen3.5-MoE 衍生，且所有基準都是小團隊自報，加權 **74.75**。 +- **Meta Muse Glimmer** 技術報告在窗口內發布，但權重 8 月就已釋出，加權 **71.5**。 +- **MemT-1.0**（arXiv 2610.26825）是窗口內的預印本，尚未同行審查，加權 **71.25**。 +- MiniMax M3.1 Flash、Google Harness2 等已在前幾天的候選中評估過，這次不重複計分。 + +依 AgentLab 規則：只有 eligible 且 ≥ 80 才寫稿，其餘候選與淘汰理由完整保留於 `data/candidates/2026-10-20.json`。 + +## 結語 + +DeepInfer 的意義在於：一家頂尖模型公司把實際撐起大規模線上服務的推論引擎、kernel 組合與部署經驗公開，並以 Apache-2.0 授權釋出。它能否被社群在其他硬體或其他模型上實際用起來，還要看後續是否補上 KV cache 與負載平衡等元件，以及第三方的實測結果。 diff --git a/src/content/missions/2026/10/2026-10-20.json b/src/content/missions/2026/10/2026-10-20.json new file mode 100644 index 0000000..56dd7b2 --- /dev/null +++ b/src/content/missions/2026/10/2026-10-20.json @@ -0,0 +1,292 @@ +{ + "schemaVersion": "1.0", + "missionId": "mission-20261020-001", + "date": "2026-10-20", + "timezone": "Asia/Taipei", + "status": "awaiting_review", + "objective": { + "goal": "Discover, score, verify and propose one outstanding AI/agent discovery for 2026-10-20 (Asia/Taipei), or document a no-publish decision.", + "sources": [ + "x", + "github", + "arxiv" + ], + "deliverables": [ + "Verified candidate batch", + "Published article (DeepInfer)", + "Mission record", + "Pull request" + ] + }, + "agent": { + "id": "grokbot", + "framework": "Grok Bot (AgentLab-bot on minim4.home)", + "model": "grok-bot", + "runId": "routine-9a1f041e-20261021T0104-minim4-daily-2026-10-20" + }, + "selection": { + "candidatesEvaluated": 24, + "selectedCandidateId": "c20261020-deepseek-deepinfer", + "selectionScore": 84.0, + "rankingArtifact": "data/candidates/2026-10-20.json" + }, + "skillsUsed": [ + { + "skill": "x-research", + "version": "1.0.0", + "status": "success", + "executionId": "exec-x-20261020-001" + }, + { + "skill": "github-discovery", + "version": "1.0.0", + "status": "success", + "executionId": "exec-gh-20261020-001" + }, + { + "skill": "arxiv-research", + "version": "1.0.0", + "status": "success", + "executionId": "exec-ax-20261020-001" + }, + { + "skill": "fact-check", + "version": "1.0.0", + "status": "success", + "executionId": "exec-fc-20261020-001" + }, + { + "skill": "article-writing", + "version": "1.0.0", + "status": "success", + "executionId": "exec-aw-20261020-001" + }, + { + "skill": "image-curation", + "version": "1.0.0", + "status": "success", + "executionId": "exec-ic-20261020-001" + }, + { + "skill": "site-publishing", + "version": "1.0.0", + "status": "success", + "executionId": "exec-sp-20261020-001" + } + ], + "timeline": [ + { + "id": "t1", + "action": "git checkout main, pull --ff-only, create branch daily/2026-10-20 (tree clean)", + "tool": "git", + "status": "success", + "startedAt": null, + "finishedAt": "2026-10-21T01:04:55+08:00", + "evidenceIds": [ + "ev-git-branch" + ] + }, + { + "id": "t2", + "action": "GitHub search (repos created in window: stars>=5 total 400, spam-heavy, plus llm/agent/mcp/model/paper/benchmark queries); releases of 30 tracked repos (21 in window); Hugging Face API scan of 40 orgs plus trending; arXiv API over HTTPS (cs.AI/LG/CL/MA/RO, submittedDate 202610191600-202610201600): totalResults 86", + "tool": "gh+python urllib", + "status": "success", + "startedAt": "2026-10-21T01:05:49+08:00", + "finishedAt": null, + "evidenceIds": [ + "ev-gh-new", + "ev-gh-releases", + "ev-hf", + "ev-arxiv" + ] + }, + { + "id": "t3", + "action": "Public X reads: 5x search_posts_all, 1x search_news, 1x get_posts_by_ids (official lab accounts, open-weights keywords, paper/repo links, launch companies, DeepInfer thread)", + "tool": "x.search_posts_all+x.search_news+x.get_posts_by_ids", + "status": "success", + "startedAt": null, + "finishedAt": null, + "evidenceIds": [ + "ev-x-search" + ] + }, + { + "id": "t4", + "action": "Verify: DeepInfer repo metadata, commits, LICENSE, README, .gitmodules, eval README and deployment Q&A; Ornith HF commits/config; Muse Glimmer HF dates and report link; Clairic blog; 5 arXiv abstracts; dedup grep over prior candidate files", + "tool": "gh+curl", + "status": "success", + "startedAt": "2026-10-21T01:07:56+08:00", + "finishedAt": null, + "evidenceIds": [ + "ev-verify", + "ev-dedup" + ] + }, + { + "id": "t5", + "action": "Write candidates JSON (24), rank with rank_candidates.py (proposed c20261020-deepseek-deepinfer 84.0), write original SVG cover and zh-TW article (status published), scaffold mission with create_mission.py", + "tool": "python", + "status": "success", + "startedAt": null, + "finishedAt": "2026-10-21T01:11:49+08:00", + "evidenceIds": [ + "ev-ranking", + "ev-article" + ] + }, + { + "id": "t6", + "action": "Run validate_content, unittest, astro check, astro build", + "tool": "python+npm", + "status": "success", + "startedAt": "2026-10-21T01:12:24+08:00", + "finishedAt": "2026-10-21T01:12:27+08:00", + "evidenceIds": [ + "ev-validate", + "ev-unittest", + "ev-astro-check", + "ev-astro-build" + ] + }, + { + "id": "t7", + "action": "Commit as agentlab-aidatatools-bot[bot], push daily/2026-10-20, open PR to main and request review from chuangtc", + "tool": "git+gh (GitHub App installation token)", + "status": "success", + "startedAt": "2026-10-21T01:13:16+08:00", + "finishedAt": "2026-10-21T01:13:22+08:00", + "evidenceIds": [ + "ev-pr" + ] + } + ], + "tests": [ + { + "name": "python scripts/validate_content.py", + "status": "passed", + "artifact": ".reports/2026-10-20/validate.log" + }, + { + "name": "python -m unittest discover -s tests -v", + "status": "passed", + "artifact": ".reports/2026-10-20/unittest.log" + }, + { + "name": "npm run check", + "status": "passed", + "artifact": ".reports/2026-10-20/astro-check.log" + }, + { + "name": "npm run build", + "status": "passed", + "artifact": ".reports/2026-10-20/astro-build.log" + } + ], + "evidence": [ + { + "id": "ev-git-branch", + "type": "git", + "ref": "branch daily/2026-10-20 from origin/main@fcf2655" + }, + { + "id": "ev-gh-new", + "type": "http_response", + "ref": ".reports/2026-10-20/gh_new_repos.json (400 repos >=5 stars created in window, dominated by game-cheat spam; AI-focused queries)" + }, + { + "id": "ev-gh-releases", + "type": "http_response", + "ref": ".reports/2026-10-20/gh_releases_window.txt (21 releases in window across 30 tracked repos)" + }, + { + "id": "ev-hf", + "type": "http_response", + "ref": ".reports/2026-10-20/hf_orgs.txt (3 new models from 40 tracked orgs in window) and ornith.txt" + }, + { + "id": "ev-arxiv", + "type": "http_response", + "ref": ".reports/2026-10-20/arxiv.xml and arxiv_entries.json (opensearch totalResults 86 for window, HTTPS)" + }, + { + "id": "ev-x-search", + "type": "api_response", + "ref": "7 public X reads; official in-window posts from DeepSeek, Ornith, Meta, OpenAI, Cloudflare, Inception, ElevenLabs, Skyfall, MiniMax, Ollama, Unsloth" + }, + { + "id": "ev-verify", + "type": "http_response", + "ref": ".reports/2026-10-20/verify_deepinfer.txt, verify_deepinfer_doc.txt, verify2.txt" + }, + { + "id": "ev-dedup", + "type": "local_command", + "ref": "grep over data/candidates/2026-10-16..19.json: no prior DeepInfer, Ornith, Glimmer, Clairic, Mercury 3 or MemT entries; MiniMax M3.1 Flash, Harness2, DiscoveryRSI already evaluated and not re-counted" + }, + { + "id": "ev-ranking", + "type": "local_command", + "ref": ".reports/2026-10-20/ranking.json (proposedCandidateId c20261020-deepseek-deepinfer 84.0, publicationAllowed true; next c20261020-x-ornith-1-5-122b 74.75)" + }, + { + "id": "ev-article", + "type": "git", + "ref": "src/content/articles/2026/10/deepseek-deepinfer-open-source-inference.md and public/media/2026/10/deepseek-deepinfer/cover.svg" + }, + { + "id": "ev-validate", + "type": "local_command", + "ref": "python scripts/validate_content.py: passed (Content OK: 4 article(s), 6 mission(s), 2 agent(s), 119 candidate(s).) @ 2026-10-21T01:12:24+08:00" + }, + { + "id": "ev-unittest", + "type": "local_command", + "ref": "python -m unittest discover -s tests -v: passed (OK) @ 2026-10-21T01:12:24+08:00" + }, + { + "id": "ev-astro-check", + "type": "local_command", + "ref": "npm run check: passed (- 0 hints) @ 2026-10-21T01:12:26+08:00" + }, + { + "id": "ev-astro-build", + "type": "local_command", + "ref": "npm run build: passed (01:12:27 [build] Complete!) @ 2026-10-21T01:12:27+08:00" + }, + { + "id": "ev-pr", + "type": "github", + "ref": "https://github.com/aidatatools/agentlab/pull/9" + } + ], + "outcome": { + "articleId": "article-20261020-001", + "githubCommit": "79b17a8ce6633d34909a93ff85b3550576c83c9d", + "githubPullRequest": "https://github.com/aidatatools/agentlab/pull/9", + "deploymentUrl": null, + "publicationStatus": "in_review" + }, + "metrics": { + "durationSeconds": 507.0, + "tokens": null, + "estimatedCostUsd": null, + "humanInterventions": 0 + }, + "retrospective": { + "successes": [ + "X, GitHub, Hugging Face and arXiv all responded; 24 deduplicated candidates scored with explicit reasons.", + "DeepSeek DeepInfer verified against the official X thread, repo commits, LICENSE, README and deployment Q&A; scored 84.0 and published." + ], + "failures": [ + "X post retrievedAt timestamps were not captured per call, so they are null.", + "The Muse Glimmer report, OpenAI low-resource languages page and MemT release artifacts were not read in full." + ], + "lessonsLearned": [ + "Repo created_at can predate public release; use first-commit time and official announcement for window checks." + ], + "nextActions": [ + "After merge and DigitalOcean deploy, run a real smoke test in a follow-up PR before promoting the mission." + ] + } +}