Research window: Past 24 hours (2026-08-13 10:19 ~ 2026-08-14 10:19 Beijing Time) Sources: GitHub (org: flagos-ai), Google News aggregation (Chinese and English), HN, BAAI Community, etc. (see appendix for details)

1. Open-Source Project Progress (GitHub Activity)

Window Overview: Of the 52 repos in the org, 16 had pushes during the window; commit search returned 46 in-window commits, with no formal Release (it has been over 7 weeks since FlagOS 2.1). Three main threads this window: FlagAttention multi-backend runtime architecture lands (8 chip backends rolled out at once), PyTorch-Plugin-FL completes org-level rename and consolidation under the name Torch-FL, and FlagGems remains high-frequency (19 commits, advancing on dual tracks: KernelGen NVIDIA operators and KMCompiler Ascend/Iluvatar operators).

1.1 FlagAttention: Multi-Backend Runtime Lands, 8 Chip Backends Rolled Out at Once

Sources: FlagAttention PR #48, FlagAttention commits

  • “add multi backends” (#48, merged Aug 13, 12:18): Adds a unified backend abstraction at src/flag_attn/runtime/backend/, establishing 8 backend packages at once—_ascend (Ascend), _enflame (Enflame), _hygon (Hygon), _iluvatar (Iluvatar), _metax and _mthreads (two MetaX lines), _nvidia. FlagAttention moves from an NVIDIA single-backend Triton attention operator library toward a FlagOS all-chip attention operator library, aligning with FlagGems’ multi-backend model. This is the most important architectural event of the window, with the backend list corresponding almost one-to-one with the chips of FlagOS member organizations.
  • Supporting commits: the three commits “add multi backends / add copy right / fix” from 08-13 02:55-04:07 (UTC) are the development-process commits for #48.

1.2 Torch-FL: Org-Level Rename and Consolidation (PyTorch-Plugin-FL → Torch-FL)

Sources: .github commits, Torch-FL commits

  • The org’s official README (English + Chinese) replaces all references to PyTorch-Plugin-FL with Torch-FL (replaced Aug 13, 18:44/18:45; links corrected 21:25): PyTorch-Plugin-FL formally enters the org mainline under the name Torch-FL (the CI badge in the Torch-FL repo README still points to the old repo name, indicating the migration is still in progress). Torch-FL is positioned as FlagOS’s PyTorch device plugin, exposing a unified flagos device and routing operators to reusable native kernels, compiler kernels, vendor-native implementations, and explicit CPU fallback.
  • Fix #99 (14:54): Preserve module parameters in non-global swap mode;
  • docs #101 (14:56): Adds a measured operator support report.

1.3 FlagGems: 19 Commits, Dual Tracks of KernelGen NVIDIA + KMCompiler Ascend/Iluvatar

Source: FlagGems commits

  • New KernelGen NVIDIA operators (Aug 13, 13:06-14:22): max_pool3d_with_indices_backward (#5264), ormqr (#5252), heaviside (#5120), leaky_relu_backward (#5145), _scaled_dot_product_attention_math (#5168), _native_batch_norm_legit_no_training (#5243); additionally, three operators—hardswish (#5139), absolute_ (#5187), arctanh (#4818)—are formally registered via “Move to ops” from the generation area—the final step in bringing KernelGen-generated operators into the formal maintenance list;
  • KMCompiler operators (Aug 14, 10:19-10:22): linalg_cross (#4850) and linalg_solve_triangular (#5201) added; Ascend pairwise_distance backend + Iluvatar CPU reference error fix (#5394)—Ascend and Iluvatar dual backends advanced in the same commit;
  • flagtree dependency formalized: #5444 Bump flagtree version for NVIDIA (15:26); #5433 updates the flagtree ascend whl to 0.6.1 in the Ascend backends.yaml (17:13)—yesterday’s judgment on the “FlagTree wheel distribution system” is confirmed: FlagTree wheels (including Ascend 0.6.1) have become a formal dependency component in FlagGems’ backend configurations;
  • Fixes and engineering: empty zero-element handling (#5438), nested_view_from_buffer_copy precision fix (#5324), xor/ixor split and registration (#5283), cat cpp wrapper path and dtype fix + rwkv_mm_sparsity precision (#5388); the fused_marlin_moe kernel is renamed per the precision scheme and its benchmark split (#5437, 10:28).

1.4 FlagGems-vllm: MetaX MC550’s mthreads Backend Enters Fused MoE

Source: FlagGems-vllm #100

  • [KMCompiler] Fused moe: add mthreads backend and tune metax MC550 (#100, Aug 13, 16:35): The Fused MoE kernel adds an mthreads (MetaX) backend and tunes for MC550—MetaX continues to advance on the operator side in vLLM scenarios (forming a compile-operator-distribution three-track complement with yesterday’s FlagTree METAX specialization unification and MACA wheel packaging);
  • v0.1.1-rc stable line: #618 removes debug logs (18:23), #679 skips some tests (19:32).

1.5 FlagCX: Allocator Compilation Process Enhancement + CI Modernization

Source: FlagCX commits

  • [Others] Enhance FlagCX allocator compilation process (#538, Aug 14, 10:03): Allocator compilation process enhancement, the only functional change in this window;
  • CI modernization: actions/checkout v4→v7 (#534), upload-artifact v4→v7 (#535), download-artifact v4→v8 (#533), [CICD] use system nlohmann-json (#530, 10:09).

1.6 vllm-plugin-FL: Enflame GCU Precision Fix

Source: vllm-plugin-FL #330

  • Fix acc mismatch for GCU (Enflame) (#330, Aug 14, 09:59): Fixes the Enflame GCU precision mismatch issue, continuing yesterday’s work on integrating the GCU300 native FlashAttention backend; no new v0.3.0 activity in this window.

1.7 build-infra: Packaging Refactor + Enflame Environment Settings

Source: build-infra commits

  • refactor: migrate wheel builders + vllm repack under packaging/ (#370, Aug 13, 15:22): Wheel builders and vllm-repack are consolidated under the packaging/ directory—release infrastructure reorganization, continuing the componentization effort after yesterday’s megatron wheel factory;
  • Update environment settings for Enflame backend (#372, 18:55): Enflame backend environment settings update.

1.8 Other Repos

  • FlagTree: #979 [AABS] Fix pruned_configs in autotuner (Aug 13, 21:45)—AABS autotuner pruning config fix;
  • FlagFFT: perf/cuda-graph branch merged into main (Aug 13, 18:39)—CUDA Graph performance optimization lands;
  • FlagPrism: New library engineering effort advances—[PROJECT] LICENSE + CODEOWNERS, [CI] code format workflow (#1, 16:35-16:55). FlagPrism (created Aug 4) is a standalone repo for FlagTree debugging/profiling components (flagtree.debugger / flagtree.profiler), consumed by FlagTree as a third_party/FlagPrism submodule; debugger/profiler wheels are no longer released separately and are merged into the FlagTree wheel;
  • docs: #460 documentation auto-sync workflow (21:31);
  • Quiet repos: FlagScale, Megatron-LM-FL, sglang-plugin-FL, FlagBLAS, etc. had pushed_at updates during the window but no in-window commits (mostly branch/tag activity, e.g., Megatron-LM-FL pushed 08-13 13:16).

II. News and Ecosystem

2.1 Alibaba Open-Sources Qwen3.8-2.4T-A95B, FlagOS Community Completes Multi-Chip Adaptation in Sync (Top Story of This Window)

Date: 2026-08-13 (17:29 Beijing time)　Source: Sina Finance (gnews)

Sina Finance reported that Alibaba open-sourced its new-generation MoE flagship model Qwen3.8-2.4T-A95B (2.4T total parameters, 95B activated), with the BAAI FlagOS community completing multi-chip adaptation in sync. On the same day (August 13, 21:25), another report noted that Qwen3.8-2.4T went live on the National Supercomputing Internet platform (caiwennews). Verification: as of the report cutoff, no explicit qwen3.8 adaptation commits were seen in the GitHub org window (commit search only returned historical qwen3-related records); the adaptation work may be in the mirroring/release preparation stage, and this item is recorded based on news reporting. This event follows the same “model open-sourcing + rapid FlagOS multi-chip adaptation” cadence as the earlier DeepSeek-V4 eight-chip Day0 adaptation.

2.2 Hygon (Member Unit) 2026 Interim Report: Revenue RMB 9.099 Billion, +66.52%; Net Profit Attributable to Parent RMB 1.798 Billion, +49.69%

Date: 2026-08-13 (from 18:45 Beijing time, multiple outlets)　Source: East Money, Sina Finance

Hygon (688041) released its 2026 interim report: revenue of RMB 9.099 billion, up 66.52% year-on-year, and net profit attributable to parent of RMB 1.798 billion, up 49.69% year-on-year, as AI computing power drives continued growth in demand for domestic high-end chips; multiple brokerages maintained or initiated buy ratings (Huaxing Securities maintained buy on a “CPU+DCU dual-engine drive,” August 13, 15:39). Media also focused on the controversy over negative operating cash flow and period-end inventory exceeding RMB 12.3 billion in “strategic stockpiling” (Sina Finance, August 14, 09:07). On the same day, a report on the “six major domestic CPU central government procurement mega-order” showed Hygon topping the list with 2,059 units (Sina Finance, August 14, 07:26). As a FlagOS member unit (Hygon DCU), Hygon’s business performance and ecosystem adaptation (FlagAttention added a hygon backend, see 1.1) are improving in tandem.

2.3 Iluvatar CoreX (Member Unit) Forecasts Turnaround to Profit in H1

Date: 2026-08-13 (17:38 Beijing time)　Source: Guandian, East Money

Iluvatar CoreX (09903.HK) issued a profit forecast: it expects to turn profitable in H1 2026, with net profit of approximately RMB 60 million to RMB 140 million (including fair value gains of up to approximately RMB 790 million). On the same day, FlagGems merged a fix for an Iluvatar CPU reference error (#5394, see 1.3) — technical adaptation and business performance are advancing in step.

2.4 Moore Threads: Dense Follow-Up Analysis on Hong Kong Listing

Date: 2026-08-13 ~ 08-14　Source: Sina Finance, CSDN

Following the launch of its H-share IPO on August 12, analytical coverage followed densely: Sina Finance published “Moore Threads: The Moment Domestic GPUs Are Closest to ‘Crossing the Snowy Mountains’” (August 14, 07:52), and CSDN published “180x Price-to-Sales! Moore Threads’ Hong Kong Listing” (09:09), while China Post Securities initiated coverage with a buy rating (August 13, 17:54). The market is focused on valuation levels and the competitive landscape for domestic GPUs. On the technical side, the Moore Threads S5000 training image pipeline landed in FlagScale yesterday (see yesterday’s report 1.1).

2.5 Overall Assessment of the News Side

Google News returned zero direct hits for component keywords such as FlagGems/FlagScale/FlagTree/FlagPerf/FlagCX/KernelGen; news in this window was dominated by model open-source ecosystem (Qwen3.8 multi-chip adaptation) and member unit business developments (Hygon interim report, Iluvatar CoreX turnaround, Moore Threads IPO follow-up), with no FlagOS-related items on HN (the C3 compiler and CoreTrace CPU simulator are both unrelated to FlagOS and were excluded). Hits from the Chinese query “BAAI open source” remained dominated by gambling/SEO noise and were excluded in full.

III. Deep Dive on Member Organizations

Multi-vendor activity map within the window (based on submissions/reports dated August 13-14):

Vendor Chip/Backend Activity within window Evidence
Moore Threads MUSA S5000 Dense post-IPO analysis (crossing the snow mountain / 180x P/S ratio / China Post buy); FlagScale S5000 pipeline was yesterday’s result Sina/CSDN/China Post Securities
Enflame Enflame GCU300 vllm-plugin-FL fixes GCU precision mismatch #330; build-infra Enflame environment setup #372; FlagAttention adds enflame backend vllm-plugin-FL/build-infra/FlagAttention
Ascend (ecosystem expansion) Ascend NPU FlagGems pairwise_distance Ascend backend #5394; flagtree ascend whl 0.6.1 #5433; FlagAttention adds ascend backend FlagGems/FlagAttention
Hygon Hygon DCU Interim revenue +66.52%, net profit +49.69%; tops central government procurement with 2,059 units; FlagAttention adds hygon backend East Money/Sina/FlagAttention
Iluvatar CoreX Iluvatar Forecasts turnaround in H1 (net profit RMB 60M-140M); FlagGems fixes Iluvatar CPU reference error #5394; FlagAttention adds iluvatar backend Viewpoint Network/FlagGems/FlagAttention
MetaX Metax GPU FlagGems-vllm Fused MoE adds mthreads backend and tunes MC550 #100; FlagAttention adds metax/mthreads backends FlagGems-vllm/FlagAttention
Horizon D-Robotics BPU No new substantive submissions/reports within window (yesterday’s Journey 7/Starry Sky already reported) —
Tsingmicro tsingmicro Nothing new within window —
KunlunXin (ecosystem) KunlunXin Nothing new within window (yesterday’s vllm-repack report already covered) —
Cambricon Cambricon Nothing new within window (news is stock price noise, not included) —

Trend assessment: The biggest signal in this window is FlagAttention rolling out 8 backends at once (ascend/enflame/hygon/iluvatar/metax/mthreads/nvidia)—the attention operator library becomes the third component after the FlagGems operator library and FlagCX communication library to achieve “full-chip coverage,” and the backend list maps closely to member organizations’ chips, proving that FlagOS multi-chip adaptation is moving from “per-chip plug-ins” to an architectural phase of “unified runtime abstraction.” Another signal is the Torch-FL renaming and consolidation: the PyTorch plug-in layer brand is unified into a single “flagos device” abstraction, which together with the FlagTree wheel merger (FlagPrism spun off independently) and build-infra packaging refactoring constitutes pre-release cleanup of “componentization + brand unification.” On the business side, Hygon and Iluvatar CoreX show improving performance, and member organizations’ fundamentals are solid.

IV. Summary

  1. FlagAttention multi-backend runtime lands: #48 merges 8 chip backend packages in one go (Ascend/Enflame/Hygon/Iluvatar/MetaX×2/NVIDIA), making the attention operator library full-chip, a key step in FlagOS’s multi-chip architecture.
  2. Torch-FL official renaming and consolidation: the org README fully replaces PyTorch-Plugin-FL with Torch-FL, and the single “flagos device” abstraction becomes the unified entry point for the PyTorch plug-in layer.
  3. FlagGems high-frequency on two fronts: KernelGen NVIDIA operators 6 new + 3 registered, KMCompiler Ascend/Iluvatar operators 3 in a row; flagtree wheel (NVIDIA bump + Ascend 0.6.1) becomes a formal dependency for each backend, fulfilling yesterday’s “wheel distribution system” judgment.
  4. MetaX MC550 continues to deepen: FlagGems-vllm Fused MoE adds mthreads backend and tunes MC550, with compilation (FlagTree METAX specialization), distribution (MACA wheel), and operators (vLLM scenarios) advancing on three fronts.
  5. News headlines: Qwen3.8-2.4T-A95B open-sourced + FlagOS community multi-chip adaptation (per Sina Finance), continuing the FlagOS ecosystem rhythm of “model open-sourcing followed by rapid multi-chip adaptation”; member organizations Hygon’s interim report and Iluvatar CoreX’s turnaround are business-side highlights.
  6. Release side: no formal Release; build-infra packaging refactoring + FlagGems-vllm v0.1.1-rc stable line, componentized version release still brewing (7+ weeks since FlagOS 2.1).

Limitations note: The Qwen3.8 multi-chip adaptation item is based on news reports; no explicit adaptation submission has yet appeared in the GitHub org window; gnews redirect links cannot fetch full text, so news details rely on headlines and multi-source cross-verification.

Appendix: Complete Source List

No. Event Source Link
1 FlagAttention multi-backend runtime #48 https://github.com/flagos-ai/FlagAttention/pull/48
2 org README renamed to Torch-FL https://github.com/flagos-ai/.github/commits
3 Torch-FL fixes/operator reports #99/#101 https://github.com/flagos-ai/Torch-FL/commits
4 FlagGems KernelGen NVIDIA operators (#5264/#5252/#5120/#5145/#5168/#5243) https://github.com/flagos-ai/FlagGems/commits
5 FlagGems KMCompiler Ascend/Iluvatar #5394/#4850/#5201 https://github.com/flagos-ai/FlagGems/commits
6 FlagGems flagtree version upgrade #5444/#5433 https://github.com/flagos-ai/FlagGems/commits
7 FlagGems-vllm mthreads MC550 #100 https://github.com/flagos-ai/FlagGems-vllm/commit/99ffbaae2244d2d3c3f99e813eedcd7a8893d5a5
8 FlagCX allocator/CI #538/#530/#533-535 https://github.com/flagos-ai/FlagCX/commits
9 vllm-plugin-FL GCU precision #330 https://github.com/flagos-ai/vllm-plugin-FL/commit/db9afd62d1257566e61f5d469de2b62844d1b379
10 build-infra packaging refactor #370, Enflame environment #372 https://github.com/flagos-ai/build-infra/commits
11 FlagTree AABS fix #979 https://github.com/flagos-ai/FlagTree/commit/7d5dbab643fd8846b88c5bb054f9953783d7be4b
12 FlagFFT cuda-graph merged https://github.com/flagos-ai/FlagFFT/commit/21235540ec852c1f6da20c5ff631653d9bba4184
13 FlagPrism LICENSE/CI https://github.com/flagos-ai/FlagPrism/commits
14 docs auto-sync #460 https://github.com/flagos-ai/docs/commit/b2808917aef14e9d4269b5c55d812707e3ad9a07
15 Alibaba open-sources Qwen3.8-2.4T-A95B + FlagOS adaptation (Sina Finance) https://news.google.com/rss/articles/CBMib0FVX3lxTE4wT1NQeHpiMGhiWkZrODYxYmdpTFFRdko2NXFqX1RPT1o2aHE1OUs5ZG5tb0pBUmZzOTFCUzdrODFtLUhjV194QnJ6Z1pmRGxiazhod2V4aUllc0VfNEZyaEFMVWVpRWJHelo1ZGd0UQ?oc=5
16 Qwen3.8-2.4T launches on Supercomputing Internet (caiwennews) https://news.google.com/rss/articles/CBMiXEFVX3lxTE42U2sybFo2cDROeTJVWE01RjR1SU1wakRmdV80bGxVcUpicHctajU1TklFV21pZ05tSXFRVmxlbmdGZExrelZRVktjR3NtdHp2WUtCa0JkYTRMeTFY?oc=5
17 Hygon 2026 interim report (East Money) https://news.google.com/rss/articles/CBMiYEFVX3lxTE5nOVlUcWUzSkdQTERMWUFUUTRvODAxcUExa1E4cG5pdmJKdmtseFdCSEdadEdMVUUzckdmMWw4VXRhOThwT0dyeDdBUm1pUHdIMjZ3WjllV1NOMFpCdUdSaw?oc=5
18 Analysis of Hygon interim report “quality” (Sina Finance) https://news.google.com/rss/articles/CBMieEFVX3lxTE12Zm9WeHE2cU1iRkRMV1czOUF0T01yZlhpZ0IxRU5Sa21ldXZwcXdDRTNuSXJIeDh4NDV0bEU3NzVNNGZ1S1M3UkxiVmdrc2lrcWwtTWJsaUoyQ01FVnpfRkpnZkdKcFU0Sy1icHp4b2ZFTEkybkh1OQ?oc=5
19 Six domestic CPU central procurement: Hygon tops (Sina Finance) https://news.google.com/rss/articles/CBMisgFBVV95cUxQQ08teEZWWjk3SDdPd3Z4bWtXR21DQ3lyMVVYNTVxRllkYlN6MWhkRUFvVjNCeEhiRTlndTYwcDd1cV9KYjBuMENBOE9LU3JqcjF2SFljY3JzdV9YNUZmbEdlTXRScHl3RGs1OEVUdU9WX2JLY0gtVEFscEU5eUkyTXJmcjE1Rlh3c29XVDBicFdiOUgyQUhHUV85azZyZ2Y4RGU0cmFGeExSVGZQbkpsNTlR?oc=5
20 Iluvatar CoreX turnaround forecast (Guandian) https://news.google.com/rss/articles/CBMiYkFVX3lxTE9KV1B4QXp3VHZKMUVObjFEN0NrcjBvUE9qYmtOSWFlbHVEdDh5SURLZm5QWEk3cWJnWFdvUndZLXNqVENfbzlDVU9WZ040LTdtdnBKNk9BdWw4cDhLemoxeE9R?oc=5
21 Moore Threads “crossing the snow mountain” (Sina Finance) https://news.google.com/rss/articles/CBMidkFVX3lxTFBxQ2tac1BMc2hpWTRNc0xOWEJLYnJ5NnlyaV9zYXBMWkVGQ044X0hXdUtRTnVLbnJ3bVo0T1o0ak9BV2JfM0htV1NlYTczWER6aEdJZTRGQlJLVzlmaUhzRjdFTXNXLVI4bnVWcUFDRVlBQ3lrVHc?oc=5
22 Moore Threads 180x price-to-sales ratio (CSDN) https://news.google.com/rss/articles/CBMia0FVX3lxTFB5SGhUNWxaN1dYOHMxVndZaWhfQ1Jta29Qb0FNRUhMQ0VLYzFTSXJyMEZJdm1tY0tMeUNHV2F6NkNjRFFMMGs3b0d6Z29WdWdvZnBHTE90aWV0bGljYndubW1fWmpTeDduV1Uw?oc=5
23 org repos overview https://api.github.com/orgs/flagos-ai/repos?per_page=100&sort=updated
24 commit search (46 within window) https://api.github.com/search/commits?q=org:flagos-ai+committer-date:%3E2026-08-13T02:19:00Z