FlagOS Daily Intelligence Report (2026-08-19)
Research window: 2026-08-18 10:19 ~ 2026-08-19 10:19 Beijing Time Sources: GitHub (org: flagos-ai), Google News aggregation (Chinese and English), HN, HuggingFace, BAAI Community, etc. (see appendix for details)
1. Open-Source Project Progress (GitHub Activity)
Window Overview: Of the 53 repos in the org, 15 had pushes during the window; commit search returned 145 in-window commits (of which FlagGems-Experimental accounted for 80, the highest single-repo single-window record in the org), with no official GitHub Release. Headline of this window: FlagGems-Experimental KernelGen shipped operators in bulk across all vendors (80 commits; multi-backend merges on the same day for NVIDIA/Hygon/Iluvatar/Moore Threads/T-Head/MetaX/Kunlunxin/Cambricon/Enflame, plus newly added Cambricon MLU C++ Runtime support and full Apache 2.0 license header backfill); also FlagTree added the SpaceMiT backend (its 14th backend), sglang-plugin-FL integrated Iluvatar and Moore Threads MUSA for the first time, and vllm-plugin-FL brought Enflame S60 into CI.
1.1 FlagGems-Experimental: KernelGen Operators Shipped in Bulk Across All Vendors (80 commits, window headline)
Source: FlagGems-Experimental commits
Merged in two waves from the afternoon to the evening of 8/18:
- First wave (8/18 17:04-17:15 Beijing, ~50 commits):
- NVIDIA side: batch of new special-function operators (special_i0, special_softmax, special_logsumexp, special_legendre_polynomial_p, special_shifted_chebyshev_polynomial_t, special_xlog1py, special_log1p, special_spherical_bessel_j0), plus basic operators such as slice and split_with_sizes.
- Chip vendor side: Hygon with 10 dedicated operators at once (cudnn_convolution, broadcast_to, adaptive_max_pool3d_backward, diff, broadcast_tensors, amp_foreach_non_finite_check_and_unscale, tile, repeat, softplus_backward, avg_pool3d_backward, nll_loss_backward); MetaX with a large batch (lgamma/lgamma_, adaptive_max_pool3d_backward, special_chebyshev_polynomial_u, alpha_dropout, linalg_cholesky, log_normal_, erfinv/erfinv_, gcd_, special_gammaln, etc.); Iluvatar (addmm_, nonzero_numpy two-pass hybrid kernel); Moore Threads (linalg_cholesky); T-Head (lcm/lcm_); Kunlunxin (digamma, arctan minimax polynomial kernel, bernoulli); ENFLAME 20260709 update.
- Engineering governance: Apache 2.0 copyright headers backfilled across all source code (#4873), batch fixes for CI timeouts/failures/OOM/precision issues (#4829), FlagTune topk lens configuration, op-monitor updates.
- Second wave (8/18 18:21-22:31 Beijing, ~30 commits): continuing the “chip-specific batch” pattern from the main repo on the morning of 8/18 — [KernelGen][Hygon] (#5557), [KernelGen][Iluvatar] (#5558), [KernelGen][MThreads] 20 operators (#5556), [KernelGen][T-Head] (#5560), and [KernelGen][MetaX] (#5559) merged in succession; Cambricon MLU C++ Runtime operator support (#5535); 13 new NVIDIA operators (compute_linear_combination, _convolution_mode, _add_relu, blackman_window, bilinear, value_selecting_reduction_backward, addmv_, resize_output, as_strided_scatter, cholesky_solve_helper, unflatten, _scaled_dot_product_efficient_attention, acos); fused_rms_norm_backward registration (#5074); KMCompiler index_copy/index_copy generic kernel optimization (#5496), log_sigmoid_backward (#5012).
Interpretation: The KernelGen operator pipeline has entered a new phase of “parallel bulk shipping across multiple vendors” — all 9 backends (NVIDIA, Hygon, Iluvatar, MThreads, T-Head, MetaX, Kunlunxin, ENFLAME, MLU) had merges within a single window, and the vendor batches (consecutively numbered #5556-#5560) indicate that each chip vendor now submits generated artifacts through a unified process. MLU support at the C++ Runtime layer and the Apache 2.0 license backfill show that while expanding operator coverage, the project is simultaneously advancing deep runtime adaptation and open-source compliance governance.
1.2 FlagGems Main Repo: KernelGen NVIDIA Operators Merged in Bulk This Morning + Multi-Chip CI Fixes (14 commits)
Source: FlagGems commits
- 8/19 09:42-09:49 Beijing, dense merge of 9 KernelGen NVIDIA operators: binary_cross_entropy (#4091), expand_copy (#4324), cumsum_ (#4365), frexp (#4433), linalg_svd (#4817), upsample_bilinear2d (#5167), _scaled_dot_product_cudnn_attention (#5182), cummin_helper (#5538, including implementation + tests + benchmark).
- 8/18: [KMCompiler][Nvidia][Ascend] linalg_lu dual-backend (#5518); [metax] operator registration fix + heuristics/tune configuration updates (#5576); Try fix Iluvatar CI job (#5441); upsample_linear1d autotune key fix to unblock metax CI (#5591); empty tensor sum fix (#5567); CI import error fix (#5573).
Interpretation: The main repo and the experimental ground are expanding in sync — the Experimental repo merges in bulk first, and the main repo follows to promote them; the dense appearance of Iluvatar/metax CI fixes shows that the multi-chip CI matrix is genuinely running and continuously troubleshooting.
1.3 FlagTree: SpaceMiT Backend Formally Merged, Backend Matrix Expands to 14 Vendors (6 commits)
Source: FlagTree commits
- #1002 (8/18 10:54 Beijing) [SpaceMiT][SPEC][CI]: Standardized the SpaceMiT backend per the FlagTree-Backend-Specialization spec into the
third_party/spacemit/specandspec_cppdirectories, with CI scripts placed underthird_party/spacemit/script, avoiding patch failures caused by main-branch directory changes. - #1004 (8/18 15:46) [DOC][CI]: README updated to include the SpaceMiT backend and workflow updated — the README changelog explicitly records 2026/08/18 addition of spacemit backend integration (based on Triton 3.6) with CI/CD.
- #999 [BACKEND][SUNRISE]: llvm dependency update for the SUNRISE backend.
- #908 [TLE]: added the tle.gpu.set_layout primitive; #826 [TLE]: tle.gpu.alloc now supports shared memory aliasing; #1003 [CI]: cleanup before ppu3.6 workflow build.
Interpretation: SpaceMiT (a RISC-V chip vendor), which previously maintained the spacemit-backend-plugin branch in its own fork, has now been formally upstreamed — FlagTree’s multi-backend matrix expands to 14 vendors (NVIDIA/TileIR/AMD/Enflame/Iluvatar/Hygon/Moore Threads/DAMO/Whale/ MetaX/SUNRISE/KLX/T-Head/SpaceMiT), adding a new member to the RISC-V camp after T-Head Zhenwu PPU.
1.4 Torch-FL: Ascend MSPTI profiler + PrivateUse1 autocast + MUSA Native RNG (5 commits)
Source: Torch-FL commits
- #126 (8/18 17:48): Added Ascend MSPTI profiler integration.
- #127 (8/18 17:48): Added PrivateUse1 backend autocast and dtype support — generalizing AMP capability for private accelerators.
- #124 (8/18 16:51): Added MUSA native RNG and FlagGems hybrid support (Moore Threads).
- #125 (8/18 15:15): Recorded Hygon FlagTree compilation verification.
- #121 (8/18 11:08): Documented permission to use upstream torch minor stable branches.
Interpretation: Multi-chip capabilities at the training plugin layer continue to be filled in; this window covers three lines — Ascend (profiler), generic AMP for private backends (PrivateUse1), and Moore Threads (RNG) — continuing the “multi-chip feature completion” main thread of 8/17-18.
1.5 build-infra: Full vllm Pipeline Integration + Ascend hw26/Cambricon Root-Cause Fixes (18 commits)
Source: build-infra commits
- vllm pipeline integration: #431 merged the three vllm wheel/plugin/app-image pipelines into a unified CI; #432 added plugin_repo input to support pre-merge fork ports; #433-#439 series of fixes (reject abbreviated SHA, NO_PROXY comment, wheel upload unlock, safe.directory, vllm-serve launcher #438, COPY path #439).
- Ascend: #429 recorded 0.24.0 CANN 8.5.0 hw26 dual-compiler E2E + CANN 9.0.0 triton path FlagGems precision; #440 recorded hw25 app-image E2E.
- Cambricon: #430 recorded the root-cause fix for cambricon index op (corresponding to flag_gems PR #5510), with the vllm-repack blacklist transitioning to an interim state.
- Other: #427 split the megatron app Containerfile into per-app training/rl files; #428 switched runtime to unzip for unpacking the triton_post_install overlay; #69fd257 refreshed the megatron-app-image workflow description (per-app image, no longer STUB).
1.6 sglang-plugin-FL: Iluvatar Backend Integrated for the First Time (4 commits)
Source: sglang-plugin-FL commits
- #34 (8/18 20:51): Added Iluvatar backend SGLang-FL operator dispatch support — Iluvatar’s first integration into the sglang plugin line.
- #54 (8/19 09:49) [e2e]: Feat/musa sglang0516 — Moore Threads MUSA end-to-end verification (sglang 0.5.16).
- #62: Engine overrides; #66: device compatibility fix.
1.7 vllm-plugin-FL: Enflame S60 Enters CI (2 commits)
Source: vllm-plugin-FL commits
- #373 (8/18 10:37) [CICD]: Added Enflame S60 CI support — echoing the “two Enflame S60 physical-machine verifications” in Torch-FL #115 on 8/17, S60 formally enters CI on the inference side.
- #382 (8/19 09:49): Fixed garbled output when using FlagGems attention.
1.8 FlagBLAS: Hygon CI Merged + L2 Benchmark Governance (12 commits)
Source: FlagBLAS commits
- #26 (8/18 18:06) hygon_ci merged (contributed by bin913): Hygon CI configuration merged; hygon multibackend markers added to hbmv/gbmv; SPR2/HPR2 support updates; axpy implementation adjustments.
- DTK 2504 attempted then reverted: fix(hygon) torch 2.4.1 (DTK 2504) + flagtree 0.6.0+hcu3.6 was merged and then promptly reverted (8/18 16:09-16:15); environment adaptation is still iterating.
- #70 (8/18 18:52) level2-operator-updates merged (AriesMTR) + #68 l2 benchmark op_name aligned with pytest markers.
1.9 Other Repos
- FlagCX #545 (8/18 12:16): [UIL] Fixed Unified IR multi-backend issues — a follow-up fix to yesterday’s #539 unifying the one-sided IR device layer.
- FlagDNN (8/18 16:34): Added the Ascend backend (commit note self-assesses “poor perf”; initial performance to be optimized) — Ascend backend enters FlagDNN for the first time.
- FlagGems-vllm: #694 disabled some tests and benchmarks, then #695 reverted (8/18 17:55-18:03); the KMCompiler path testing strategy is being adjusted.
- FlagTensor: A v0.2.0-rc2.post1 tag appeared (a non-default branch push on 8/18 23:11); subsequent 0.2.x versions are still in preparation.
- Megatron-LM-FL: The 0.2.0-rc2 branch and branches such as copilot/fix-unit-tests and copilot/research-triton-kernel-analysis are active (pushed 8/18 20:25); the 0.2.0 series is in preparation.
- docs: Model list auto-sync branch updated; release-info: gh-pages only refreshed the build-infra deployment page (not a version signal).
- Quiet repos: FlagScale, FlagPerf, FlagAttention, FlagQuantum, FlagOS-Robo, KernelGen, whisper-triton, etc. had no pushes during the window.
II. News Coverage and Ecosystem
2.1 Overall Assessment of News Coverage
Google News returned zero direct hits within the window for the component keywords FlagGems/FlagScale/FlagTree/FlagPerf/FlagCX/KernelGen; the HN query yielded only one false match (“Clinch – Warp fork,” unrelated to FlagOS, excluded). The “BAAI open source” query remained dominated by gambling/SEO noise and generic AI aggregation posts from the BAAI community (all excluded in bulk). Component-level news was quiet, with effective ecosystem items concentrated in member organizations’ commercial developments, detailed in 2.3 and Chapter III.
2.2 Ecosystem Reference Items (within the 8/18 window)
| Date | Event | Source |
|---|---|---|
| 8/18 | UCloud public cloud officially launches Hygon Tianxi series: Hygon DCU servers deployed at scale on a public cloud for the first time (multi-source coverage) | Sina Finance |
| 8/18 | Hygon C86/DCU dual-chip solution written into national roadmap (Drivers Home) | Drivers Home |
| 8/18 | CAICT Institute of AI Research, together with Hygon and industry organizations, releases a research report on enterprise computing power foundation construction | Sina Video |
| 8/18 | Moore Threads plans Hong Kong listing, as semiconductor firms ride the wave to Hong Kong | Sina Finance |
| 8/18 | Interview with MetaX’s Peng Li: chips iterate every “two to three years,” models every “two to three months” — how can domestic GPUs keep up (SEMI semiconductor industry site) | ecar.semi.org.cn |
| 8/18 | Iluvatar CoreX: RMB 700M in book gains props up turnaround, GPU main business still awaiting “positive turn” (business-side analysis) | thebambooworks.com |
| 8/18 | Cambricon: first batch of equity incentives for 124 employees unlocked, average book gain of about RMB 5.57M per person (business-side, multi-source coverage) | EE Times China |
| 8/18 | BiSheng compiler core AscendNPU IR open-sourced, Huawei architect Hai Lijuan details technical progress, supports Triton and other operator ecosystems (BAAI community report, echoing the FlagTree-AscendNPU-IR repo within the org) | BAAI Community |
| 8/18 | BAAI open-sources AquilaMed-RL medical reasoning RL model (HuggingFace BAAI org updated 8/18, along with the CCI3-HQ-Classifier data tool) | https://huggingface.co/BAAI/AquilaMed-RL |
Excluded items: Pinzhun Laser IPO, Unitree listing, Zhipu GLM-5.3 Hong Kong stock fluctuations, STAR Market ETF heavyweight stock bulletins, Cambricon share-price financial flashes, and other items with no substantive connection to FlagOS; broad chip geopolitics topics such as “lend me some compute resources” are treated as background only.
2.3 Trends over the Past 3 Days (corroborated by GitHub org pushed activity)
| Date | Active repos | Main content |
|---|---|---|
| 8/17 | FlagGems(25), Torch-FL(10), build-infra(16), FlagBLAS(7), FlagCX, FlagFFT, Megatron-LM-FL, FlagTree, and 16 repos in total | Wave of batch KernelGen operator merges; Torch-FL multi-chip feature completion (DCU AMP, Enflame torch_gcu-free); Hygon BW1000 added to CI |
| 8/18 | FlagGems-Experimental(80), build-infra(18), FlagGems(14), FlagBLAS(12), FlagTree(6), Torch-FL(5), sglang-plugin-FL, vllm-plugin-FL, and 15 repos in total | KernelGen ships in bulk across all vendors (9 backends merged the same day); FlagTree adds SpacemiT backend; sglang/vllm plugins go multi-chip; build-infra vllm pipeline integration |
| 8/19 (tail of this window) | FlagGems, sglang-plugin-FL, vllm-plugin-FL | FlagGems main repo merges 9 KernelGen NVIDIA operators in the morning; sglang MUSA e2e; vllm-plugin garbled-text fix |
Trend assessment: The “multi-vendor parallel rollout” pace of 8/17–18 accelerated further in this window — the KernelGen pipeline achieved full coverage of 9 backends in a single window, and shifted from “filling in operator coverage” to “vendor-specific batches + deep C++ Runtime adaptation”; the FlagTree backend matrix expanded to 14 vendors and welcomed a new RISC-V member, SpacemiT; on the inference side (sglang/vllm plugins), Iluvatar CoreX, Moore Threads, and Enflame S60 backends and CI began to be filled in. It has been 8 weeks since FlagOS 2.1 (6/24), with component feature volume and validation records continuing to accumulate; FlagTensor 0.2.x (rc2.post1 tag) and the Megatron-LM-FL 0.2.0-rc2 branch suggest a new round of releases may be imminent.
III. Deep Dive into Member Organizations
Multi-vendor activity map within the window:
| Vendor | Chip/Backend | Activity within the window | Evidence |
|---|---|---|---|
| BAAI | — | FlagCX UIL multi-backend fixes (#545); build-infra vllm full-pipeline integration and multi-vendor validation records; open-sourced AquilaMed-RL medical RL model | FlagCX #545 / build-infra #431-440 / HF BAAI |
| Hygon | Hygon DCU | FlagGems-Experimental 10+ dedicated operators in a single window (#5557 batch, etc.); FlagBLAS hygon_ci merged (#26) + DTK 2504 adaptation attempt; Torch-FL records FlagTree compile validation (#125); commercial side: UCloud Tianxi public cloud launch, C86/DCU written into national-level roadmap, CAICT computing power foundation report | FlagGems-Exp #5557 / FlagBLAS #26 / Torch-FL #125 / gnews multiple sources |
| Enflame | Enflame GCU | vllm-plugin-FL Enflame S60 enters CI (#373); FlagGems-Experimental ENFLAME 20260709 update | vllm-plugin-FL #373 / FlagGems-Exp |
| Moore Threads | MUSA | sglang-plugin-FL MUSA e2e (#54); Torch-FL MUSA native RNG (#124); FlagGems-Experimental MThreads 20-operator batch (#5556) + linalg_cholesky; commercial side: planning Hong Kong IPO | sglang-plugin-FL #54 / Torch-FL #124 / FlagGems-Exp #5556 / gnews |
| Iluvatar | Iluvatar | sglang-plugin-FL Iluvatar backend integrated for the first time (#34); FlagGems-Experimental Iluvatar batch (#5558) + nonzero_numpy optimization; FlagGems main repo Iluvatar CI fix (#5441); commercial side: turnaround analysis coverage | sglang-plugin-FL #34 / FlagGems-Exp #5558 / FlagGems #5441 |
| MetaX | MetaX | FlagGems-Experimental MetaX large batch (#5559 and 15+ operators); FlagGems main repo metax registration fix (#5576) + upsample_linear1d autotune fix unblocks CI (#5591); commercial side: Peng Li interview on iteration cadence | FlagGems-Exp #5559 / FlagGems #5576/#5591 / gnews |
| Cambricon | Cambricon MLU | FlagGems-Experimental MLU C++ Runtime operator support (#5535); build-infra records index op root-cause fix, vllm-repack blacklist moved to transitional state (#430); commercial side: equity incentive unlock | FlagGems-Exp #5535 / build-infra #430 / gnews |
| Ascend (non-member of the ecosystem) | Ascend | Torch-FL Ascend MSPTI profiler (#126); FlagDNN adds Ascend backend; FlagGems linalg_lu Ascend backend (#5518); build-infra CANN 8.5.0 hw26 dual-compiler E2E (#429) + hw25 records (#440); BiSheng AscendNPU IR open-sourced (ecosystem) | Torch-FL #126 / FlagDNN / FlagGems #5518 / build-infra #429 / gnews |
| KunlunXin | KunlunXin | FlagGems-Experimental Kunlunxin 3 operators (digamma, arctan minimax, bernoulli) | FlagGems-Exp #288-290 |
| T-Head | T-Head Zhenwu PPU | FlagGems-Experimental T-Head batch (#5560); FlagTree ppu3.6 CI cleanup (#1003) | FlagGems-Exp #5560 / FlagTree #1003 |
| SpacemiT (new entrant) | SpaceMiT | FlagTree SpaceMiT backend formally merged (#1002/#1004, recorded in README changelog on 8/18) | FlagTree #1002/#1004 |
| SUNRISE | SUNRISE | FlagTree SUNRISE backend llvm dependency update (#999) | FlagTree #999 |
| Tsingmicro / Horizon | TXDA / BPU | No open-source activity within the window | — |
Trend assessment: The chip landscape in this window is characterized by “full-vendorization of the KernelGen pipeline + outbound expansion of the backend ecosystem.” On one hand, 9 backends shipped simultaneously within a single window, and vendors have shifted from “being adapted” to “self-submitting via a unified process” (evidenced by the consecutive vendor batches #5556-#5560); on the other hand, FlagTree welcomed new RISC-V member SpacemiT, the inference plugin lines (sglang/vllm) began multi-chip expansion, and Ascend is flourishing across all four fronts of Torch-FL/FlagDNN/FlagGems/build-infra — FlagOS’s “unified software stack” is permeating a broader range of chips and frameworks.
IV. Summary
- FlagGems-Experimental logged 80 commits in a single window, with KernelGen shipping in batches across all vendors (headline of this window): 9 backends merged on the same or next day, including vendor-specific batches (#5556-#5560), Cambricon MLU C++ Runtime support, the NVIDIA special-function family and 13 new operators, plus full Apache 2.0 license annotation — the operator pipeline has entered a phase of “multi-vendor parallelism and deep adaptation.”
- FlagTree backend matrix expands to 14: The SpaceMiT backend was formally upstreamed per spec with CI/CD added, adding another member to the RISC-V camp after T-Head Zhenwu PPU; the SUNRISE backend was updated in tandem.
- Multi-chip acceleration of inference/training plugins: sglang-plugin-FL integrated the Iluvatar backend for the first time and completed Moore Threads MUSA e2e; vllm-plugin-FL Enflame S60 entered CI; Torch-FL added the Ascend MSPTI profiler and generic PrivateUse1 autocast.
- build-infra mechanism upgrade: The three-stage vllm wheel/plugin/app-image pipeline was unified and integrated (#431); after the Cambricon index op root-cause fix, the vllm-repack blacklist moved to a transitional state; Ascend hw26 dual-compiler E2E records.
- Member organizations active on the commercial front: Moore Threads planning a Hong Kong IPO, Hygon Tianxi series public cloud scaled launch, Iluvatar turnaround analysis, Cambricon equity incentive unlock — software stack adaptation and capital market developments are advancing in tandem.
- Quiet on the news side: zero hits on component keywords, no valid HN results, no official GitHub Release; the FlagTensor v0.2.0-rc2.post1 tag and the Megatron-LM-FL 0.2.0-rc2 branch suggest a new round of releases may be imminent.
Limitations: FlagGems-Experimental commit counts and merge times come from commit search (the search index may lag slightly behind actual merges; individual commit times are based on the repos API pushed_at); the “poor perf” conclusion for the FlagDNN Ascend backend comes from commit comments, with no independent benchmark data seen; gnews redirect links cannot be scraped for full text, so news-side judgments rely on headlines and cross-verification from multiple sources; SUNRISE and SpaceMiT are inferences as new members of the FlagTree backend matrix, and neither has yet appeared on the official org member list.
Appendix: Complete Source List
| No. | Event | Source Link |
|---|---|---|
| 1 | FlagGems-Experimental KernelGen batch merge of operators across all vendors (~80 entries, including vendor batches #5556-#5560, MLU C++ Runtime #5535, Apache 2.0 annotation #4873) | https://github.com/flagos-ai/FlagGems-Experimental/commits/main |
| 2 | FlagGems main repo KernelGen NVIDIA 9 operators (#4091/#4324/#4365/#4433/#4817/#5167/#5182/#5538) + linalg_lu dual backend (#5518) + metax/Iluvatar CI fixes | https://github.com/flagos-ai/FlagGems/commits/main |
| 3 | FlagTree SpaceMiT backend merge (#1002/#1004), SUNRISE llvm (#999), TLE set_layout/alloc aliases (#908/#826) | https://github.com/flagos-ai/FlagTree/commits/main |
| 4 | Torch-FL Ascend MSPTI profiler (#126), PrivateUse1 autocast (#127), MUSA RNG (#124), Hygon FlagTree validation (#125) | https://github.com/flagos-ai/Torch-FL/commits/main |
| 5 | build-infra vllm pipeline integration (#431-#439), Ascend hw26 E2E (#429/#440), Cambricon root-cause fix (#430) | https://github.com/flagos-ai/build-infra/commits/main |
| 6 | sglang-plugin-FL Iluvatar backend (#34), MUSA e2e (#54) | https://github.com/flagos-ai/sglang-plugin-FL/commits/main |
| 7 | vllm-plugin-FL Enflame S60 CI (#373), flaggems attention fix (#382) | https://github.com/flagos-ai/vllm-plugin-FL/commits/main |
| 8 | FlagBLAS hygon_ci (#26), L2 benchmark governance (#68/#70) | https://github.com/flagos-ai/FlagBLAS/commits/main |
| 9 | FlagCX UIL multi-backend fix (#545) | https://github.com/flagos-ai/FlagCX/pull/545 |
| 10 | FlagDNN Ascend backend | https://github.com/flagos-ai/FlagDNN/commits/main |
| 11 | FlagTensor v0.2.0-rc2.post1 tag / Megatron-LM-FL 0.2.0-rc2 branch | https://github.com/flagos-ai/FlagTensor/tags |
| 12 | org repos overview (53 repos) | https://api.github.com/orgs/flagos-ai/repos?per_page=100&sort=updated |
| 13 | commit search (145 entries within window) | https://api.github.com/search/commits?q=org:flagos-ai+committer-date:%3E2026-08-18T02:19:00Z |
| 14 | Google News Chinese and English component queries | https://news.google.com/rss/search?q=FlagOS+when:14d |
| 15 | HN Algolia | https://hn.algolia.com/api/v1/search_by_date?query=FlagOS |
| 16 | UCloud launches Hygon Tianxi series (8/18) | https://news.google.com/rss/articles/CBMieEFVX3lxTE1DODdrY29TS1I5NTlGcVlILVUyTC1OUnNndU9oR2FFRGpIRmFTQUhfN1V4S2p0MTFtdDJDRW5Pb2JWUzJ2OVhfOFZFbjlqNFQzYXlrZGdmamxfa0ZHYWl5MnhNV1VJNXdpazVJMFpDVEZYUURVNmtySw?oc=5 |
| 17 | Hygon C86/DCU dual-chip solution written into national roadmap (8/18) | https://news.google.com/rss/articles/CBMiWEFVX3lxTE9Razk5Z3ljbWVrOXNXOW9nSWRjOUdQM0FYYldnS283dWl1Qmhod01GY2c4Wmp0TGlUZTZTMUNMSjhGNWc5VXJIa3VOd1ZmemV1SVVtX2g1RGU?oc=5 |
| 18 | CAICT and Hygon jointly release enterprise computing foundation report (8/18) | https://news.google.com/rss/articles/CBMiggFBVV95cUxPcVFVX2dPcG96cVdmYVBOVXNaaXh0dlU1TXJQU0N4N2ZPRTQxQ0hNdkEwZklZLU1NRThZVDNZQ2x0dGd2c1l6ajNVTVJ0R2cxZFBFT3RaVF9HOTlDbGNPVHB6d2lGSjBVMkdxWUZVRjVhTG5mYjNwRVlUZzRKV1Fn?oc=5 |
| 19 | Moore Threads plans Hong Kong listing (8/18) | https://news.google.com/rss/articles/CBMidkFVX3lxTE5velk2S05Tb1M5R1MwdGFsZzNXZjNEaWRkWk9WdzI0SHplYm90eFNYUGJuLTNTY2N4VWFiMGtQNjd4S09ISE9ZZWhYTERNdWV1QzNfSEQ2V1lsM0duUzg1NGpIYUMtNGQxWVFCZzQzZGVScDlscHc?oc=5 |
| 20 | Interview with MetaX Peng Li (8/18, SEMI) | https://news.google.com/rss/articles/CBMihgFBVV95cUxNM3B1Zk1ORG1pR3EwMG5DU3NUSzVDZURqWXFPajctanNJX2F4V2xSMVl6Z3MwLVI4U1VkZFNjNUYzNlBqaXE1WUMzYnJUOUdBcXFEZVRZLUktNGxqZi1qMDR5VTVNMVdMWW84a3VaSGtIR0ZqNW5zTUthRVNHOGNQY1ltal9Ndw?oc=5 |
| 21 | Iluvatar CoreX turnaround analysis (8/18) | https://news.google.com/rss/articles/CBMiuwJBVV95cUxOYV80YnVwSHBsaTJmMm5PQjM4OXIzNmc3bjR5VkN3T2xmanlMRDZRZERaM0hpNDJIWUVZcmpJZmlDZENxYlpsRG1PMUZPd2lXZmxvWlpQN2NiU2FEeWxNdUN1M2dxU3R6UmowdE9laGpJVTNYRTNxVzI4ajYxYjNHZjFBbVhDYklYT1hrSV80T0tWZ0JuZ2NrR3VwRDQyOUc5OVp1cS1Cc2RNRkVMdTJKbDN5dmRxM1ZGamVyemhwamJDcjdqWjdnMGUxMUEtQ1hpNDhraFJjdXdHajRIZnVVR2R3aVhGYjk3S0dUSjhrc3R3QkNINTBjOXg4d2otOWtfQzBYVzVfcHNCOWlhamhwYjc2b1hWbmhLRjhiOHlsc0piMnJhT3l6NlBxMXpBVmMydnhZa2VoNDBIOFU?oc=5 |
| 22 | Cambricon equity incentive unlock (8/18) | https://news.google.com/rss/articles/CBMiU0FVX3lxTE9YVzY3d2g3Qlo2aE9fWDVmRVc3OVFyb3pUSHo2Zzh5TFlRVHYwbTJxZ0J4UTEya1ZfcEN2cjIyUkVKeFlJdmJPSXg0UVE0RzF1bjdr?oc=5 |
| 23 | Bisheng compiler AscendNPU IR open-sourced (8/18, BAAI Community) | https://news.google.com/rss/articles/CBMiSEFVX3lxTFA0bUNEUkdPeXgtSnVkYnVIc05PY1JSQXV4d1dpckRrSHRSSXduYmlCVFdxdm1SM0RqdTlGdkt0UGY1OU0yT1NjZw?oc=5 |
| 24 | BAAI AquilaMed-RL (HuggingFace) | https://huggingface.co/BAAI/AquilaMed-RL |