Research window: 2026-08-26 10:18 ~ 2026-08-27 10:18 Beijing Time Sources: GitHub (org: flagos-ai 52 repos + commit search + Issues/PR/Release API), Google News RSS (multiple queries in Chinese and English, proxy chain operational), HN Algolia, Tavily cross-validation (see appendix for details)

Index

    1. Open-Source Project Progress (GitHub Activity)
      • 1.1 Moore Threads Megatron dual-backend verification loop closed — musa4.3.6/5.2.0 full pipeline turns green (window headline) (08-26/08-27)
      • 1.2 vllm-plugin-FL: Hygon backend and Qwen3.6 W4A16 promoted to main line, release/0.2 line maintains MLU590/Ascend/Kunlunxin (08-26)
      • 1.3 build-infra: verl-FL app image build plan, sglang packaging research, continued verification engineering hardening (08-26/08-27)
      • 1.4 FlagGems: community PR surge — SiliconFlow joins for the first time, Moore Threads/Iluvatar optimizations, new KernelGen operators, Cambricon bugfix (08-26/08-27)
      • 1.5 FlagGems-vllm: Qwen4 in-house Triton kernels and vendor baselines (08-26)
      • 1.6 Torch-FL: Ascend torch.compile landing, Hygon DCU low-precision matrix software path, routing rollback (08-26)
      • 1.7 Other repository activity (FlagCX, FlagTree, Megatron-LM-FL, sglang-plugin-FL, FlagTensor, FlagBLAS, FlagDNN, docs) (08-26/08-27)
    1. News Coverage and Ecosystem
      • 2.1 Component-level keywords hit zero for the eighth consecutive day
      • 2.2 Embodied intelligence ecosystem: Second World Humanoid Robot Games conclude (08-25/08-27)
      • 2.3 Ecosystem exclusions and upcoming milestones
    1. Member Organization Deep Dive
      • 3.1 BAAI: Peking University–BAAI joint team wins table tennis championship at robot games (08-25/08-26)
      • 3.2 Tsingmicro: member organization engineer reports operator registration bug in FlagGems (08-26)
      • 3.3 Hygon: post-listing follow-up — vllm-plugin-FL backend and Qwen3.6 W4A16, Torch-FL low-precision path (08-26)
    1. Summary
  • Appendix: Complete Source List

I. Open Source Project Progress (GitHub Activity)

Window Overview: Of the 52 repos in the org, 17 had pushes during the window; commit search returned 82 in-window commits (build-infra 57, FlagGems 6, FlagGems-vllm 3, FlagCX 3, Torch-FL 3, FlagTree 2, FlagBLAS 2, FlagTensor 2, docs 2, sglang-plugin-FL 1, FlagDNN 1, Megatron-LM-FL 1), plus several repos with PR branch pushes (vllm-plugin-FL, flagtree-cpu, community, release-info, FlagGems-Experimental). Main threads this window: 1) Inference-side framework integration and multi-backend maintenance continue to deepen — vllm-plugin-FL’s Hygon backend and Qwen3.6 W4A16 move onto the main line, the release/0.2 maintenance line adds multi-backend PRs for Cambricon MLU590/Ascend/Kunlunxin, and build-infra adds a verl-FL app image build plan plus sglang packaging research; 2) Moore Threads Megatron dual-backend verification loop closes — the full Megatron-LM-FL training/post-training/inference verification matrix turns green on both musa4.3.6 and musa5.2.0, making it one of the most complete backends on the training side of the verification matrix; 3) torch.compile coverage extends to Ascend, with Hygon DCU following on the low-precision path. The news side remains quiet, consistent with the past seven days, while GitHub stays hot.

1.1 Moore Threads Megatron Dual-Backend Verification Loop Closes — musa4.3.6/5.2.0 Full Pipeline Turns Green (Window Headline)

Sources: build-infra #573, #577, #578, #579, #589, #590

  • musa5.2.0 backend: 8/26 21:34 fixed the issue that “Moore Threads E2E requires --distributed-backend mccl” (#573); 22:14 marked megatron training as verified (#577); 22:51 enabled the megatron_training app image config for the mthreads backend (#578); 23:05 recorded app image tag 2.1.2-0.2.1_14_g6f61e9c5b; 23:10 marked post_training/inference verification passed (F+T dual path) (#579).
  • musa4.3.6 backend: 8/27 09:46 recorded the same app image tag; 09:48 marked training/post_training/inference all verified (#590); 09:50 merged.
  • MetaX side same day: 8/26 17:31 recorded and merged the result of 1 megatron verification cell for metax-maca3.7.2.1 (#549).

Interpretation: Moore Threads becomes the member in the verification matrix whose Megatron-LM-FL training-side “training–post-training–inference” full pipeline is verified on both MUSA versions (4.3.6 and 5.2.0), with training-side coverage more complete than any prior member. Combined with yesterday’s Hygon DCU three-line entry, the verification matrix is moving from “inference side turning green first” to “deep training-side coverage,” and the templated replication effect of the verification paradigm continues to show. The musa dual-backend F/T dual path (vllm side) reported on 8/24 and this megatron-side closure corroborate each other.

1.2 vllm-plugin-FL: Hygon Backend and Qwen3.6 W4A16 Move to the Main Line, release/0.2 Line Maintains Multiple Backends (08-26)

Sources: #407, #411, #409, #408, #406, #392

  • Main line: #407 (created 8/26 10:49) adds the Hygon backend with Qwen3.6 W4A16 support (Vendor PR, another step forward for Hygon DCU on the inference side); #406 consolidates the Iluvatar triton patch to module level and cleans up dead code.
  • release/0.2 maintenance line (v0.2 series wrap-up; #405 already switched target branch yesterday): #411 (8/26 19:39) adds Cambricon MLU590 support, verified end-to-end on both neuware 4.7.2 and 4.4.3 stacks (5 touchpoints: device registration, vendor identification, etc.); #409 Ascend generic full-decode graph mode; #408 Kunlunxin P800 CI support; #392 uploads benchmark serve reports to FlagCICD.
  • Kunlunxin decode scale fix #400 remains open (created 8/24; continues as yesterday’s watch item).

Interpretation: vllm-plugin-FL shows a dual-track rhythm of “main line opening a new backend (Hygon) + 0.2 maintenance line wrapping up while continuously adding backends (MLU590/Ascend/Kunlunxin).” Hygon completed HCU vendor work in sglang-plugin-FL yesterday and today submitted the Hygon backend and Qwen3.6 W4A16 in vllm-plugin-FL — Hygon’s coverage on the FlagOS inference side is moving from “entry” to “multi-framework, multi-precision.”

1.3 build-infra: verl-FL App Image Build Plan, sglang Packaging Research, Verification Engineering Hardening Continues (08-26/08-27)

Sources: #588, #582, #572, #575, #574, #567, #571

  • verl-FL app image build plan (#588, merged 8/27 09:37): the RL framework verl-FL (the FlagOS fork of ByteDance’s RL framework verl) enters app image build planning — the third training/inference framework to join the image line after vllm/megatron.
  • sglang release-packaging research (#582, submitted 8/26 23:52): adds a research report on sglang release packaging, laying groundwork for the sglang-plugin-FL release process.
  • Cambricon neuware4.4.3 enables vllm 0.20.2 app dep (#567): feat branch merged, followed by recording app image tags 2.1.2-0.2.1_gef78dec.d20260826 and 2.1.2-0.2.1_gb954912.d20260826 (8/26 18:15, 23:05) — Cambricon’s dual software stacks (4.7.2/4.4.3) verification lines advance in parallel.
  • Verification engineering hardening continues: skip runtime installation in app-image mode (#572); explicit --stack-version prevents stale node configs from silently driving image tags (#575); verify-vllm-backend.sh unifies a fixed-path repo root (#575); vllm verification reports split from “one giant document” into per-version directories (#574), with narrative tables rewritten per markdown conventions and temporary node names removed (privacy/readability cleanup); docs explicitly state the “upstream PRs require review” principle (#571).
  • Image line: base/runtime image descriptions refreshed for 2.1.2 (#565/#566); dependabot batch-upgrades GitHub Actions (checkout 4→7, setup-python 5→7, upload-artifact 4→7, download-artifact 4→8, dawidd6 23→24).

Interpretation: build-infra’s substantive moves this window are “expanding framework integration surface (verl plan, sglang research) + engineering verification credibility (explicit versions, structured reports, privacy cleanup).” If verl-FL completes app image integration, FlagOS will cover four mainstream framework categories: pre-training (megatron), inference (vllm), RL (verl), and inference serving (sglang), completing the last category in its framework map.

1.4 FlagGems: Community PR Wave — SiliconFlow’s First Entry, Moore Threads/Iluvatar Optimizations, New KernelGen Operators, Cambricon Bugfixes (08-26/08-27)

Sources: FlagGems commits (6 merged in-window) + open PRs (14 newly created in-window)

  • Merged to main: two Cambricon bugfixes — “index autotune no longer kills the process on untileable expand_shape” (#5510), “paged load_from_kvcache broadcasts scalar max_virtual_index” (#5786); Iluvatar python version bump (#5765); KMCompiler Ascend cholesky solve layout test stabilization (#5698); QC FP8 W8A8 bmm optimization (#4880); new PR validation rule-check CI workflow (#5774).
  • Newly created PRs (community side): SiliconFlow’s first contribution — CI test aliases and approved operator list (#5788), new special_laguerre_polynomial_l operator (#5782); three Moore Threads optimizations (conv_transpose1d stride-phase gather #5789, matmuladd reusing addmm SQMMA/FMA kernels #5781, upsample_linear1d_backward #5779); Iluvatar restores rank-1 vectorization under Triton 3.6 (#5783); four new KernelGen Nvidia operators (native_norm #5770, fused_moving_avg_obs_fake_quant #5772, sspaddmm #5773, nested_to_padded_tensor #5775); Hopper MM dispatch and tuning optimization (#5780); cleanup of dead ATen registrations and non-standard registration names (#5777); index generation cache corruption recovery (#5784).

Interpretation: FlagGems this window shows three tracks advancing together: “vendor PRs + independent developer PRs + KernelGen operator generation.” SiliconFlow, a compute service provider that is not a member organization, contributing code directly to the operator library is a positive signal of FlagOS’s community “co-build” character; Moore Threads/Iluvatar continue to feed the operator library via PRs, complementing build-infra’s verification line.

1.5 FlagGems-vllm: Qwen4 In-House Triton Kernels and Vendor Baselines (08-26)

Sources: #713, #681, #709

  • #713 (merged 8/26 22:02) adds Qwen4 in-house Triton kernels and vendor baseline comparisons — preparing in-house kernels for the new-generation model and incorporating vendor-native kernels as performance baselines.
  • #681 renames the fused_marlin_moe kernel by precision scheme and splits the benchmark; #709 narrows the fail-closed scope of select tests.

Interpretation: Qwen4 kernels entering FlagGems-vllm shows the operator library’s model-following cadence has advanced from Qwen3.5/3.8 to the Qwen4 series; the “in-house Triton kernels + vendor baselines” comparison mode makes performance conclusions quantifiable.

1.6 Torch-FL: Ascend torch.compile Lands, Hygon DCU Low-Precision Matrix Software Path, Routing Rollback (08-26)

Sources: #189, #208, #196

  • #189 (8/26 11:04) lands Ascend torch.compile support via triton-ascend — following 8/25’s Enflame GCU torch.compile (#187), the torch.compile sequence now covers four backend categories: MetaX/MUSA/GCU/Ascend.
  • #208 (8/26 19:24) adds a DCU low-precision matrix software path — Hygon DCU’s low-precision (low-bit) matrix computation software path in Torch-FL, in sync with vllm-plugin-FL’s Hygon Qwen3.6 W4A16.
  • #196 rolls back Postroggy’s FlagGems routing-related PR (8/26 13:32); reason remains to be seen (FlagGems also had fix PRs submitted by Postroggy on 8/24-25).

1.7 Other Repository Activity (08-26/08-27)

  • FlagCX: [CICD] openEuler 24.03 joins the NVIDIA RPM build matrix (#549, 8/27 09:24); PAL layer SHMEM adaptor accepts opaque init context handles (#557).
  • FlagTree: [VERSION] export triton_key to align with the Triton 3.4 compiler interface (#1052, 8/26 19:50); TLE fixes “default to (1,1,1) when the backend does not declare cluster_dims” (#1020).
  • Megatron-LM-FL: adds chunked cross-entropy to reduce memory usage (#126, 8/26 12:58) — memory optimization for long-sequence training.
  • sglang-plugin-FL: [e2e] adds the T-Head PPU (PPU-ZW810E) CI platform (#64, 8/26 10:47) — T-Head PPU inference-side verification is formally integrated.
  • FlagTensor: fixes Iluvatar vendor-native baseline resolution after the main merge (#16, 8/26 17:48).
  • FlagBLAS / FlagDNN: CI installs nvidia-cuda128 extras fix (#94 merged + follow-ups); Ascend performance fixes.
  • docs: model list synced with ModelScope (#469) and HuggingFace (#470, 8/26 13:54).

II. News Coverage and Ecosystem

2.1 Component-level keywords see zero hits for the eighth consecutive day

Across multiple gnews queries in Chinese and English (FlagOS/FlagGems/FlagScale/FlagTree/FlagPerf/FlagAttention/FlagCX/KernelGen/FlagOS-Robo/FlagQuantum/FlagPrism/FlagBLAS when:14d + 智源研究院 开源 when:7d + BAAI open source when:7d), exact component-name hits within the window were zero; HN Algolia queries (FlagOS/FlagGems/FlagScale/FlagTree/FlagPerf/KernelGen/BAAI) returned only irrelevant entries; Tavily cross-validation found no new component-level FlagOS coverage. The pattern of “quiet on the news side, red-hot on the GitHub side” has now held for eight consecutive days, and the official v0.3.0 release remains the next news trigger.

2.2 Embodied intelligence ecosystem: Second World Humanoid Robot Games conclude (08-25/08-27)

Sources: Beijing Daily Sports/Beijing Daily Online (8/25), ITHome (8/25), Sing Tao Headline via gnews (8/27)

  • The Second World Humanoid Robot Games (8/22-26, “Ice Ribbon”) concluded. On 8/25, in the 100m semifinals, Tiangong from the Beijing Humanoid Robot Innovation Center ran 8.86 seconds to break the record again (ITHome); on 8/27, Sing Tao Headline reported that Tiangong won the title in 8.64 seconds, breaking the 100m record for a third time.
  • Evening of 8/25, table tennis final: the joint team of Peking University and BAAI won the championship (see 3.1).
  • The event featured 666 teams, 2,056 robots, and 51 competition categories; results in multiple events surpassed human records, and events such as table tennis were contested “fully autonomously,” testing AI decision-making rather than remote control.

Interpretation: The event provides a concentrated stress-testing ground for embodied intelligence. On the FlagOS side, the corresponding FlagOS-Robo embodied toolchain (whose end-to-end data pipeline was tracked in the August report) has not yet generated component-level news in this window, but the ecosystem buzz of “validating products through competition” is a positive backdrop for FlagOS’s embodied direction.

2.3 Ecosystem-side exclusions and upcoming milestones

  • A signed article on People’s Daily Online, “International Observation: Consensus Implemented to Benefit the World — Interpretation of the 2026 World Artificial Intelligence Conference Chairman’s Statement” (reposted by BAAI Community on 8/27) — signed by BAAI scholars, a policy-interpretation piece rather than FlagOS component developments, so not included in the main text;
  • BAAI Community’s “ChinaTalk The ‘Bubble’ of Chinese AI” think-tank observation (8/26) — macro commentary;
  • “87.4 Billion, the World’s Largest AI Open Source Community Is Up for Sale” (follow-up on the HuggingFace acquisition rumor) — unrelated to FlagOS;
  • Upcoming milestone: FlagOS official (CSDN account, 8/12) announced a call for papers for the OASIS-AI workshop (co-located with MICRO 2026, 11/1, Athens, Greece, focused on open software ecosystems for heterogeneous AI systems and FlagOS multi-chip adaptation practices), with a submission deadline of 8/31 — 4 days to the deadline, worth watching for submissions and subsequent agenda disclosures.

III. Member Deep Dive

3.1 BAAI: PKU-BAAI Joint Team Wins Table Tennis Championship at Robot Games (08-25/08-26)

Sources: Beijing Daily Sports/Beijing Daily Online (8/25), China News Service via gnews (8/26)

  • On the evening of 8/25, the table tennis final of the 2nd World Humanoid Robot Games (a newly added event, with over ten teams) concluded. The Peking University-BAAI Joint Team won the championship with a “stability first” strategy (steady returns, waiting for opponent errors), defeating the University of Hong Kong SMASH team 2-1 in a best-of-three format.
  • To level the hardware playing field, the competition uniformly used the AgiBot Expedition A3 humanoid robot body, with teams competing on AI compute, model training capability, trajectory prediction, and whole-body motion control—the PKU-BAAI Joint Team deployed multiple match strategies (serve strategy, spin strategy) and plans to continue evolving in both spin and footwork.
  • China News Service reported the event at 11:18 on 8/26 (within the window).

Analysis: BAAI’s embodied intelligence efforts (joint team with PKU, running AI models on robot bodies) echo the direction of FlagOS’s FlagOS-Robo toolchain. Tasks like table tennis—requiring “millisecond-level reaction and autonomous decision-making”—impose extremely high demands on on-device inference latency, making them a typical scenario for edge inference and embodied models.

3.2 Tsingmicro: Member Unit Engineer Reports Operator Registration Bug in FlagGems (08-26)

Source: FlagGems #5752 (8/26 10:06)

  • The official Tsingmicro account tsingmicro-public-e filed an issue in FlagGems: the registration for special_i0e in __init__.py should be written as ("special_i0e.out", special_i0e_out) (pointing to the line in the Iluvatar commit merged at 8/26 09:41)—a member unit engineer directly participating in operator library code review and issue reporting.

Analysis: Tsingmicro’s involvement in FlagOS has extended from “backend adaptation” to “co-building operator library code quality”—a micro-level signal of deepening member unit engagement.

3.3 Hygon: Follow-up to Onboarding—vllm-plugin-FL Backend with Qwen3.6 W4A16, Torch-FL Low-Precision Path (08-26)

Sources: vllm-plugin-FL #407 (8/26 10:49), Torch-FL #208 (8/26 19:24)

  • Following 8/25’s “Hygon three-line onboarding” (FlagScale BW1000 CI, FlagCX DCU adaptation, sglang-plugin-FL HCU vendor), Hygon advanced two more steps in this window: vllm-plugin-FL submitted the Hygon backend with Qwen3.6 W4A16 support (created 8/26, main line); Torch-FL added a DCU low-precision matrix software path (merged 8/26).
  • Hygon now covers four lines on the FlagOS side: training (FlagScale CI), communication (FlagCX), inference (sglang-plugin-FL + vllm-plugin-FL), and framework layer (Torch-FL), while deepening into low precision (W4A16).

IV. Summary

Window Main Threads: 1. Inference-side framework integration and multi-backend maintenance deepening—vllm-plugin-FL Hygon backend + Qwen3.6 W4A16 promoted to main line; release/0.2 line adds Cambricon MLU590, Ascend full-decode graph mode, and Kunlunxin P800 CI; build-infra adds a verl-FL app image build plan (the framework landscape will soon include RL) and sglang packaging research. 2. Moore Threads Megatron dual-backend verification loop closed—the full training/post-training/inference verification matrix for musa4.3.6 and musa5.2.0 all turned green, with training-side coverage ranking first in the verification matrix. 3. torch.compile sequence covers Ascend (GCU→Ascend, four backends now complete), with Hygon DCU low-precision path following up (W4A16 + low-precision matrix). 4. Community co-building spillover—SiliconFlow contributes operators to FlagGems for the first time; Moore Threads, Iluvatar CoreX, and Tsingmicro participate in operator library co-building via PRs/issues. 5. Member units: the PKU-BAAI Joint Team wins the table tennis championship at the Robot Games, with BAAI’s embodied direction echoing FlagOS-Robo.

Next Watch Points: vllm-plugin-FL v0.3.0 official release (rc0 is out, Qwen3.5 text-only compatibility #383 already merged to main); merge status of Kunlunxin decode scale #400 and P800 CI #408; first verification results after Hygon #407 is merged; verl-FL app image landing pace; OASIS-AI call for papers deadline 8/31; FlagGems 5.3.5 full-backend re-acceptance wrap-up; follow-up handling of Torch-FL #196 rollback.

Limitations: Commit counts and merge times come from commit search and the repos/commits API (search indexing may lag slightly; pushes within the window for repos such as vllm-plugin-FL, flagtree-cpu, community, and release-info were PR branch pushes with no new merges to the main branch, verified using the pulls/issues API); “SiliconFlow’s first contribution” is based on that org account first appearing in FlagGems commit history (no full org history check was performed); gnews redirect links cannot be scraped for body text, so news-side judgments rely on headlines and cross-verification across multiple sources; Robot Games results are per media reports.

Appendix: Complete Source List

No. Event Source Link
1 build-infra mthreads E2E mccl fix (#573) https://github.com/flagos-ai/build-infra/pull/573
2 build-infra mthreads-musa5.2.0 megatron training verified (#577) https://github.com/flagos-ai/build-infra/pull/577
3 build-infra mthreads megatron_training app image enabled (#578) https://github.com/flagos-ai/build-infra/pull/578
4 build-infra mthreads-musa5.2.0 F+T verified (#579) https://github.com/flagos-ai/build-infra/pull/579
5 build-infra mthreads-musa4.3.6 full pipeline verified (#589/#590) https://github.com/flagos-ai/build-infra/pull/590
6 build-infra metax-maca3.7.2.1 megatron results (#549) https://github.com/flagos-ai/build-infra/pull/549
7 vllm-plugin-FL Hygon + Qwen3.6 W4A16 (#407) https://github.com/flagos-ai/vllm-plugin-FL/pull/407
8 vllm-plugin-FL Cambricon MLU590 + neuware dual stack (#411) https://github.com/flagos-ai/vllm-plugin-FL/pull/411
9 vllm-plugin-FL Ascend full-decode graph (#409) https://github.com/flagos-ai/vllm-plugin-FL/pull/409
10 vllm-plugin-FL Kunlunxin P800 CI (#408) https://github.com/flagos-ai/vllm-plugin-FL/pull/408
11 vllm-plugin-FL Iluvatar triton patch integration (#406) https://github.com/flagos-ai/vllm-plugin-FL/pull/406
12 build-infra verl-FL app image build plan (#588) https://github.com/flagos-ai/build-infra/pull/588
13 build-infra sglang release-packaging investigation (#582) https://github.com/flagos-ai/build-infra/pull/582
14 build-infra Cambricon neuware4.4.3 vllm 0.20.2 (#567) https://github.com/flagos-ai/build-infra/pull/567
15 build-infra verify hardening (#572/#575) https://github.com/flagos-ai/build-infra/pull/575
16 build-infra vllm report split and cleanup (#574) https://github.com/flagos-ai/build-infra/pull/574
17 FlagGems Cambricon index autotune fix (#5510) https://github.com/flagos-ai/FlagGems/pull/5510
18 FlagGems Cambricon KV cache fix (#5786) https://github.com/flagos-ai/FlagGems/pull/5786
19 FlagGems SiliconFlow CI alias (#5788) https://github.com/flagos-ai/FlagGems/pull/5788
20 FlagGems Moore Threads conv_transpose1d optimization (#5789) https://github.com/flagos-ai/FlagGems/pull/5789
21 FlagGems Iluvatar rank-1 vectorization restored (#5783) https://github.com/flagos-ai/FlagGems/pull/5783
22 FlagGems KernelGen native_norm (#5770) https://github.com/flagos-ai/FlagGems/pull/5770
23 FlagGems-vllm Qwen4 kernels and vendor baselines (#713) https://github.com/flagos-ai/FlagGems-vllm/pull/713
24 Torch-FL Ascend torch.compile (#189) https://github.com/flagos-ai/Torch-FL/pull/189
25 Torch-FL DCU low-precision matrix path (#208) https://github.com/flagos-ai/Torch-FL/pull/208
26 Torch-FL Postroggy routing rollback (#196) https://github.com/flagos-ai/Torch-FL/pull/196
27 FlagCX openEuler 24.03 RPM matrix (#549) https://github.com/flagos-ai/FlagCX/pull/549
28 FlagTree triton_key aligned with Triton 3.4 (#1052) https://github.com/flagos-ai/FlagTree/pull/1052
29 Megatron-LM-FL chunked cross-entropy (#126) https://github.com/flagos-ai/Megatron-LM-FL/pull/126
30 sglang-plugin-FL T-Head PPU-ZW810E CI (#64) https://github.com/flagos-ai/sglang-plugin-FL/pull/64
31 FlagGems Tsingmicro issue (#5752) https://github.com/flagos-ai/FlagGems/issues/5752
32 org repos overview (52 repos, 17 active within window) https://api.github.com/orgs/flagos-ai/repos?per_page=100&sort=updated
33 commit search (82 unique commits within window) https://api.github.com/search/commits?q=org:flagos-ai+committer-date:%3E2026-08-26T02:18:00Z
34 Peking University & BAAI team wins table tennis championship (Beijing Daily Sports / Beijing Daily Online) https://news.bjd.com.cn/2026/08/25/11932197.shtml
35 Table tennis championship report (China News Service via gnews) https://news.google.com/rss/articles/CBMifkFVX3lxTE11UExBLWFCdUpaVzFZM3VFRXh1MHJRU0lhVWhFMDFTM2dzODc2MHFsYjZrdTFlOTJRRVhTMldNbkNGYmZDdC13ZHdNVUpPTFM0SzBPTFJpenRGa2UzZy1kbFlWREt0c0thVhXNUUNocHFQOUswcHYyRkJ4a2x6UQ?oc=5
36 Tiangong 100m in 8.86 seconds (ITHome) https://www.ithome.com/0/994/182.htm
37 Tiangong breaks record three times with 8.64 seconds (Sing Tao Headline via gnews) https://news.google.com/rss/articles/CBMirAJBVV95cUxNRmh3VlZLaW5Tbloyb3NSWXRiRFRhUllfQUFQNXh3WG5zaGtTZU1ObGN0TGxxZ0U5MWx5U3JkUTZfV0V1X0VwSy1ybUFRLUdHNzdhM3IxdXFBbG1OY1ZxQmIzSzdOWEE1dWE0SzV6bjlUVlFwVFNoZXB6c25oYUpudVFWVFgxR1ZBbmJFWGF6UUx5a3N0TkQ3YTJlS2RIVGFCb09OYVBCN1c1UGVxc2dUeHhZX3dtYVdVRnFfckVvU1R0aW0tVTFQaGR3UmF1OVAxcGk0a3pnRmFPNWRZSDJHY2UzbXd6SkkwbTRJX3RCMDFlVUhZM01xaFhnTHNDNW1KMWRwQ29HZ0VNV1hGX2cwMXg0TUg4Y1k5R1RYaHEwWXF2a09sdkZCbkUwUTg?oc=5
38 OASIS-AI call for papers (FlagOS CSDN official account) https://flagos.csdn.net
39 gnews multi-query in Chinese and English (zero hits at component level) https://news.google.com/rss/search?q=FlagOS+when%3A14d&hl=zh-CN&gl=CN&ceid=CN%3Azh-Hans
40 HN Algolia (irrelevant false matches) https://hn.algolia.com/api/v1/search_by_date?query=FlagOS
41 Tavily cross-validation (BAAI/member unit news) https://www.tavily.com