[agentlab-aidatatools-bot]

Daily research 2026-10-20 (Asia/Taipei)

Window: UTC 2026-10-19T16:00:00Z – 2026-10-20T16:00:00Z. Run on minim4.home by GrokBot (routine-9a1f041e-20261021T0104-minim4-daily-2026-10-20).

Candidates (24 after dedup)

Source	Candidates with this source
X (7 public reads)	13
GitHub (new-repo search + 30 tracked repos, 21 in-window releases)	9
arXiv (API over HTTPS, totalResults 86)	5
Hugging Face (40 orgs + trending)	2

Already evaluated on earlier days and not re-counted: MiniMax M3.1 Flash, Google Harness2, DiscoveryRSI.

Top 5 scores

Score	Candidate	Decision
84.0	DeepSeek DeepInfer open-source inference framework (Apache-2.0)	eligible, published
74.75	Ornith-1.5-122B open weights (MIT)	rejected: < 80, HF weights uploaded 10-16, self-reported benchmarks
71.5	Meta Muse Glimmer technical report	rejected: < 80, weights released 2026-08-09
71.25	MemT-1.0 (arXiv 2610.26825)	rejected: < 80, preprint
68.25	Halcyon T2I without language pretraining (arXiv 2610.26863)	rejected: < 80

Verification (DeepInfer)

• Official @deepseek_ai thread 1/4–4/4 at 2026-10-20T09:30:19Z (in window).
• GitHub API: repo created 2026-10-16T07:47:23Z; first public commit d585e65 "initial commit" 2026-10-20T04:09:42Z; HEAD 774ce24; Apache-2.0 LICENSE; README, .gitmodules, eval/README.md and doc/inference_deployment_qa.md read.
• Limits stated in the article: only DeepSeek-V4.1-Flash, only NVIDIA SM100 / Ascend 950, KV cache + E/P/D disaggregation + load balancing not included, all performance/scale numbers self-reported, code not run.

Files

• data/candidates/2026-10-20.json
• src/content/articles/2026/10/deepseek-deepinfer-open-source-inference.md (status: published, zh-TW)
• public/media/2026/10/deepseek-deepinfer/cover.svg (original illustration)
• src/content/missions/2026/10/2026-10-20.json (awaiting_review, no deploy claimed)

Tests (run on minim4 in .venv)

• python scripts/validate_content.py: passed (Content OK: 4 article(s), 6 mission(s), 2 agent(s), 119 candidate(s))
• python -m unittest discover -s tests -v: passed (6 tests, OK)
• npm run check: passed (0 errors, 0 warnings, 0 hints)
• npm run build: passed (16 pages, includes /discoveries/deepseek-deepinfer-open-source-inference/)

Merging this PR is the human review step. @chuangtc please check the article on DEV before approving. Not merged by the agent.

[chuangtc]

Ok