Research window: 2026-09-03 10:18 ~ 2026-09-04 10:18 Beijing Time Sources: GitHub (org: flagos-ai 52 repos + full commit search of 51 entries + per-repo supplementary checks + tags/releases verification), Google News RSS (30+ query sets in Chinese and English, proxy链路正常), HN Algolia, BAAI Community (see appendix)

Index

  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 build-infra: Enflame vLLM 0.24.0 delivery line rolled back and re-registered — tops1.10.6 verified and marked delivered (09-03/09-04)
    • 1.2 build-infra: sglang application image records landed — MetaX metax first registration, Ascend switches to CANN 9.0.0 (09-03)
    • 1.3 FlagGems-Experimental: KernelGen Moore Threads specialized operators batch-expanded by 19 in four hours (09-03)
    • 1.4 FlagTree: ROCm volatile semantics preserved, Moore Threads TLE fix, copyright header update (09-03/09-04)
    • 1.5 FlagGems-vllm: KMCompiler converges vendor backends, DeepSeek-v3.2 grouped_topk optimization, Hopper w8a8 fp8 (09-03/09-04)
    • 1.6 Other components: FlagGems fixes and CI cleanup, TransformerEngine-FL Adam interface completion, FlagCX/FlagDNN/Torch-FL/sglang-plugin-FL (09-03/09-04)
  • II. News Coverage and Ecosystem
    • 2.1 News side sees third consecutive quiet window: zero direct component-level hits, release lull continues (09-03~09-04)
    • 2.2 BAAI institutional news: First cohort of “AI+ Application and Governance Pilot Program” successfully concludes (09-03)
  • III. Member Company Deep Dive
    • 3.1 Enflame: Allotment numbers released (20,657), RMB 6.1 billion fundraising destination and “Tencent dependence” become media focus (09-03/09-04)
    • 3.2 Moore Threads: MUSA ecosystem developers surpass 800,000, claims full-chain embodied intelligence training and deployment (09-03)
    • 3.3 Hygon: Wins China Patent Gold Award; xKylin completes adaptation for 12 international industrial software products (09-03)
    • 3.4 MetaX: Non-recurring losses extend to four and a half years, lock-up expiry test approaches; appears at space computing forum (09-03)
    • 3.5 Iluvatar CoreX and industry perspective: Hong Kong stocks up over 8%, JPMorgan lists as top pick; “Four Little Dragons” interim cash flow pressure becomes consensus (09-03)
  • IV. Summary
  • Appendix: Complete Source List

1. Open-Source Project Progress (GitHub Activity)

Window Overview: Of the 52 repos in the org, 14 had pushes during the window. Commit search returned 51 in-window commits (build-infra 17, FlagGems-Experimental 19, FlagGems 4, FlagTree 3, TransformerEngine-FL 3, FlagGems-vllm 3, sglang-plugin-FL 2, etc.). Tags/releases API verification shows no new tags and no new releases in the window (the latest tags/releases for active repos were all created before the window: vllm-plugin-FL v0.3.0-rc0 released 08-24, FlagGems v5.3.0 released 06-24, etc.), continuing the stable release phase of 2.2. Main threads this window: 1) build-infra “first reverts, then re-registers, then marks delivered” the Enflame vLLM 0.24.0 delivery line — yesterday’s registered tops1.9.10 delivery record was rolled back (#703), then the same source fingerprint (g7c6758c) was used to separately verify two toolchain versions, tops1.9.10 and tops1.10.6, and re-register them, finally marking “delivered (F+T verified)”; 2) sglang application image records land for the first time for MetaX metax and Ascend CANN 9.0.0, a clear signal of an overall Ascend toolchain upgrade (the same day FlagGems CI removed cann850); 3) FlagGems-Experimental’s KernelGen Moore Threads (MThreads) specialized operators land 19 in a four-hour batch, combined with FlagTree’s TLE fix for mthreads3.6, Moore Threads adaptation enters a “dense expansion of specialized operators” phase; 4) FlagGems-vllm’s KMCompiler side optimizes grouped_topk performance for DeepSeek-v3.2 and aligns the generic Fused MoE operator with upstream, converging platform work onto vendor backends.

1.1 build-infra: Enflame vLLM 0.24.0 delivery line rolled back and re-registered — tops1.10.6 verified then marked delivered (09-03/09-04)

Sources: build-infra #703, #702, #701, #715 (commit records see build-infra commits), #716, #717, #700 (commit records same as above)

  • 09-03 12:25 #703 merged: revert the vLLM 0.24.0 delivery record registered yesterday for Enflame enflame — the tops1.9.10 application image tag registered yesterday at 22:39 (#698) was rolled back entirely, indicating that when delivery verification fails, the record is revoked first; in engineering discipline, “records are reversible and traceable.”
  • Supporting actions: 12:18 #702 removed the “distrust flagtree” wording regarding the Enflame image from the vLLM docs (the compatibility concern between the Enflame side and FlagTree was resolved); 12:18 #701 changed the app-image verify completion criterion to be based on a semantic anchor, no longer relying on easily misjudged textual signals.
  • Re-registration and delivery: 09-03 23:58 #715 registered enflame-tops1.9.10 with source fingerprint g7c6758c (d20260903); 09-04 00:02 #716 registered enflame-tops1.10.6 with the same fingerprint — the Enflame toolchain was verified from 1.9.10 to 1.10.6, with image tags strictly one-to-one with source commits; 09-04 00:23 #717 marked enflame 0.24.0 as delivered (7c6758c, F+T verified).
  • MetaX registration on the same image line: 09-03 11:30 #700 recorded the vllm application image tag for metax-maca3.8.1.3 (2.1.2-0.2.1_g5c511da.d20260901).

Interpretation: This is the first complete closed loop in build-infra of “an entire delivery registration being rolled back, then re-verified, then delivered” — after yesterday’s (09-02) Enflame tops1.9.10 registration, today it was first reverted, then separately verified with the dual toolchain versions tops1.9.10/tops1.10.6, and finally delivered with 1.10.6, indicating that the Enflame 0.24.0 application image deliverable underwent a substantive change within one day (toolchain upgrade or re-release after verification failure); #717’s “F+T verified” marker solidifies “verification complete” as a precondition for delivery. The registration-verification-delivery mechanism for the multi-vendor image matrix is now fully operational, and this is also the first member delivery in the 2.2 cycle to complete this closed loop end to end.

1.2 build-infra: sglang application image records land — MetaX metax first registration, Ascend switches to CANN 9.0.0 (09-03)

Sources: build-infra #699, #704, #706, #707, #710, #711, #712 (commit records see build-infra commits), #713, #714, #709

  • sglang application image records land for the first time: 09-03 21:46 #712 recorded the sglang application image tag for metax-maca3.8.1.3 (MetaX) (2.1.2-0.1.dev1_g0900c244b); 23:00 #713 recorded the sglang application image tag for ascend-cann9.0.0 (Ascend) (2.1.2-0.1.dev1_g2e568482e) — the Ascend toolchain version jumped directly from CANN 8.5.0 to 9.0.0.
  • Dense runtime fixes (Phase 5 wrap-up): 11:31 #699 filled sglang app-image runtime gaps (shim input, CT pinning, verify fixes); 14:29 #704 consumed the renamed shim artifact sgl-kernel-shim; 15:53 #706 pinned scipy<1.18 in application and verify installs; 16:05 #707 set PYTHONPATH=/opt/triton on every app-layer pip install; 17:23 #710 switched models to be provided via –model-path, supported app-image model input, and added Ascend CT dependencies.
  • Workflow mechanization: 21:09 #711 added a record_only mode to the sglang app-image workflow (register only, no build, aligned with yesterday’s vllm-side #696); 17:23 #709 changed the vllm plugin commit to “derived + asserted,” prohibiting manual entry; 23:56 #714 cleaned up docs for completed tasks.

Interpretation: The sglang verification matrix (yesterday Phase 5 landed the 0.5.18 application image) entered its “recording and runtime fixes” tail phase this window: sglang image tags for MetaX and Ascend began to be registered in batches, with Ascend registering directly with CANN 9.0.0 (yesterday’s vllm line was still cann8.5.0), corroborating the same-day FlagGems CI “removal of cann850” (see 1.6) — the Ascend toolchain is being switched to 9.0.0 across the board. Mechanisms such as record_only and derived+asserted indicate that build-infra has gradually replaced “manual image registration” with “CI-derived assertions + replayable records,” continuously reinforcing the engineering foundation of the multi-vendor image matrix within the 2.2 cycle.

1.3 FlagGems-Experimental: KernelGen Moore Threads specialized operators expand by 19 in a four-hour batch (09-03)

Source: FlagGems-Experimental commits (#291~#309)

  • 09-03 12:51~16:48 19 [KernelGen][MThreads] Moore Threads specialized operators merged consecutively within four hours: square_, deg2rad_, clip_, silu_backward, log2, fix, lcm, rad2deg, arctan_, masked_scale, mish_, index_copy, adaptive_max_pool3d_backward, amp_foreach_non_finite_check_and_unscale_, diagonal_scatter, channel_shuffle, log10_, feature_dropout, functional_sym_constrain_range_for_size (PR #291~#309 range).
  • Operator coverage: elementwise math ops (square_/log2/log10/rad2deg_/arctan_/deg2rad_/clip_/fix/lcm/mish_), normalization and masking (masked_scale/feature_dropout), indexing and tensor structure (index_copy/diagonal_scatter/channel_shuffle), pooling and training aids (adaptive_max_pool3d_backward/amp_foreach_non_finite_check_and_unscale_/silu_backward), etc., covering both inference and training paths.

Interpretation: This is the largest single-day batch of operator merges for Moore Threads (MThreads) within the FlagOS ecosystem — all 19 specialized operators went through the KernelGen generation path, indicating that the Moore Threads side has entered a pipeline state of “using KernelGen to batch-fill MUSA specialized operators” (the Day0 article’s “8 new operators” reported on 09-01~09-02 was the starting point of this cadence). Specialized operators proceeding in parallel with generic operators and merging at a cadence of 20~30 minutes each shows that operator expansion has become highly templated. This echoes Moore Threads’ official statement the same day that “the full chain for embodied intelligence training and deployment has been opened up” (see 3.2) — training-side operators (such as amp_foreach_non_finite_check_and_unscale_, adaptive_max_pool3d_backward) are direct evidence of filling gaps in the training pipeline.

Sources: FlagTree #1092, #1094, #1095

  • 09-03 23:26 #1092: preserve volatile loads in AMD/ROCm LLVM lowering — fixes compiler optimizations breaking volatile semantics (same theme as the ROCm adaptation line reported on 09-01~09-02, with AMD-side correctness continuing to be polished).
  • 09-03 23:31 #1094: fix the tle_supported list for TLE on mthreads3.6 — correction of TLE (Tensor Library Extension) capability detection for Moore Threads toolchain version 3.6, same origin and cadence as the MThreads operator batch expansion in 1.3.
  • 09-04 02:53 #1095: repo-wide Copyright headers updated to the FlagOS copyright notice (routine maintenance).

Interpretation: The three commits in the FlagTree window span two backend lines, AMD/ROCm and Moore Threads, and are all fixes at the “compiler path correctness” level; #1094 directly serves mthreads3.6 — Moore Threads is advancing both operator and compiler lines on the FlagOS side on the same day (1.3 + 1.4), indicating that its adaptation has reached deep into the toolchain rather than remaining only at the operator API layer.

1.5 FlagGems-vllm: KMCompiler converges vendor backends, DeepSeek-v3.2 grouped_topk optimization, Hopper w8a8 fp8 (09-03/09-04)

Sources: FlagGems-vllm #693, KMCompiler Fused MoE commit, #92

  • 09-03 14:17 #693: optimize w8a8 block fp8 matmul on Hopper — FP8 quantized matrix multiplication optimization for the vLLM inference hot path (Nvidia Hopper platform).
  • 09-04 10:07 KMCompiler Fused MoE: generic operators synchronized and aligned with FlagGems upstream, platform-specific work converged into vendor-specific backends — a “generic/vendor” layered refactor of Fused MoE: the generic part follows FlagGems upstream evolution, while vendor differences (MThreads/Iluvatar, etc.) are collected into their respective backend directories.
  • 09-04 10:09 #92: optimize grouped_topk performance for DeepSeek-v3.2 — performance optimization on the KMCompiler side for the DeepSeek-v3.2 sparse MoE routing operator (grouped_topk), directly serving domestic-chip inference for the DeepSeek series models.

Interpretation: FlagGems-vllm is the bridge repo between vLLM and FlagGems/KMCompiler. The three commits in the window represent three lines respectively: Nvidia-side FP8 performance (#693), the “generic alignment + vendor convergence” layering of the MoE architecture (Fused MoE commit), and operator performance for running DeepSeek-v3.2 on domestic chips (#92) — the core FlagOS inference chain of “DeepSeek model + KMCompiler generation + domestic backend” continues to be optimized, and grouped_topk is a widely recognized hotspot operator for DeepSeek-series MoE inference.

1.6 Other components: FlagGems fixes and CI cleanup, TransformerEngine-FL Adam interface completion, FlagCX/FlagDNN/Torch-FL/sglang-plugin-FL (09-03/09-04)

Sources: FlagGems #4128, #5948, #5963, #5970, TransformerEngine-FL #108/#109/#110, FlagCX #566, sglang-plugin-FL #43, Torch-FL #258, FlagDNN commit

  • FlagGems: 10:34 #4128 fixed cumprod precision and the non-in-place integer path issue on the SiliconFlow backend (SiliconFlow appearing as a new backend in operator library fix records is worth noting); 15:30 #5948 changed setup.sh to install python-build-standalone from a mirror site (domestic network adaptation); 20:11 #5963 extended the use_gems check to all test files changed in PRs (CI coverage hardening); 09-04 09:47 #5970 CI removes cann850 (Ascend CANN 8.5.0 exits CI, corroborating the CANN 9.0.0 registration in 1.2).
  • TransformerEngine-FL: 09-03 18:39 (10:39 UTC) three consecutive commits #108/#109/#110 completed the Adam optimizer interface for the FlagOS backend, completed the Adam interface for the Reference backend, and aligned reference compute-scale semantics — the training-stack plugin layer pins the optimizer behavior of the “reference implementation” and the “FlagOS implementation” to the same semantics (regression test baseline).
  • FlagCX: 15:29 #566 added txda_stream support to the UIL adapter (the cross-chip communication library’s UIL layer extends a new stream type).
  • sglang-plugin-FL: 13:56 #43 added support for empty-device scenarios, targeting national platform (xinchuang) deployments; 14:38 #83 fixed the container_volumes path in CUDA CI.
  • Torch-FL: 12:23 #258 flagos backend supports uint16/uint32/uint64 dtype conversion.
  • FlagDNN: 11:26 added cpu_op tests.

Interpretation: The in-window actions of long-tail components show a “platformization wrap-up” characteristic — FlagGems’ SiliconFlow backend fix suggests the operator library’s adaptation surface is still expanding (cloud vendor side); CANN 8.5.0 being removed from CI and 9.0.0 being registered in build-infra is the synchronized landing of the Ascend toolchain switch across the full FlagOS stack (operator library + image management); TransformerEngine-FL aligning the Adam semantics of the reference implementation and the FlagOS implementation is quality infrastructure for “multi-implementation cross-verification.”

II. News Coverage and Ecosystem

2.1 Third Consecutive Quiet Window on the News Side: Zero Direct Component-Level Hits, Release Plateau Continues (09-03 to 09-04)

Source: 30+ Google News RSS queries in Chinese and English (see appendix)

  • In this window (09-03 10:18 to 09-04 10:18), all Chinese and English queries for FlagOS and its core components (FlagGems/FlagScale/FlagTree/FlagPerf/FlagAttention/FlagCX/FlagOS-Robo/FlagQuantum) returned zero hits; “BAAI open source” and “FlagOS when:3d” were also empty; the two HN Algolia hits were both noise (a YC co-founder match on a security warning, and a Zenodo link challenge post) and were excluded. This is the third consecutive window with zero direct component-level hits (since 09-02).
  • Trend over the past 3 days (corroborated by GitHub): After the dense 2.2 RC0 validation tag period from 8/31 to 9/1, 9/2 to 9/4 entered a “release plateau + engineering wrap-up” phase — no new tags, no new releases, but build-infra completed the Enflame vLLM 0.24.0 delivery loop (1.1), the sglang image matrix record landed (1.2), and FlagGems-Experimental Moore Threads operator batch expansion continued (1.3). The next public action on the 2.2 checklist remains either an rc0.postN iteration or the formal rc branch cut, signaled by activity in the community/FlagRelease repository (in this window, FlagRelease had a push on 09-01 and release-info had a push at 09-04 00:03, both non-default-branch actions with no substantive version information).
  • Chinese queries still contain reprints of yesterday’s Day0 articles on third-party channels and a large volume of Enflame IPO financial content (see Chapter III). The ecosystem narrative is still digesting the multi-chip adaptation results of “three Day0s in four days,” with no new official release to follow up.

2.2 BAAI Institutional News: First Cohort of the “AI+ Applications and Governance Leadership Program” Successfully Concludes (09-03)

Source: BAAI Community, “First Cohort of the ‘AI+’ Applications and Governance Leadership Program Successfully Concludes” (BAAI Community, 09-03 23:30)

  • On the evening of 09-03, the BAAI Community published “First Cohort of the ‘AI+’ Applications and Governance Leadership Program Successfully Concludes”: the first cohort of the training/leadership program organized by BAAI in the “AI+ applications and governance” direction has wrapped up — like yesterday’s report on the conclusion of the SCO-framework AI training program, this is a routine institutional development on BAAI’s AI talent cultivation and international cooperation side, with no direct connection to the FlagOS technology ecosystem.
  • The BAAI Community remains highly active on the content side (reposts of the Gemini 3.8 release preview, Yanshuicheng’s team’s Scaling Harness Intelligence, embodied intelligence series articles, etc.), but these are all third-party content aggregations, not BAAI’s own achievement releases; there were no new official BAAI model/platform announcements within the window.

III. Deep Dive into Member Units

3.1 Enflame: Allotment Numbers Announced (20,657), RMB 6.1B Raise Allocation and “Tencent Dependence” Become Media Focus (09-03/09-04)

Sources: Securities Times (allotment numbers), East Money (20,657 allotment numbers), EE Times China (RMB 6.1B raise allocation), The Paper (issue market cap RMB 61.1B), China.com Finance (chairman responds to Tencent dependence), Cailian Press (IPO by the numbers)

  • IPO process advances: From 11:00 on 09-03, Enflame Technology (688801.SH) batch-announced its online allotment numbers — a total of 20,657 allotment numbers, each entitling the holder to subscribe for 500 shares, dovetailing with yesterday’s reported final allotment rate of 0.02455315% and 42.066 billion valid subscription shares; the RMB 6.119 billion issue size, ~RMB 61.1 billion issue market cap, and 81% three-year revenue CAGR form the backdrop of the “new share subscription” narrative.
  • EE Times China long-form article at 14:09 on 09-03, “Enflame Technology Is Also Coming to IPO — Where Will the RMB 6.1B Raise Go”: breaks down the use of proceeds — after deducting issuance expenses, funds go to R&D of next-generation AI accelerator cards/training cluster products, full-stack software and ecosystem building, and working capital supplementation (per Enflame’s prospectus), with the software/ecosystem investment share being a point of interest from the FlagOS perspective.
  • The “Tencent dependence” issue goes public: China.com Finance and others report the chairman’s response on the eve of listing — Tencent is both Enflame’s largest customer and largest external shareholder (the strategic investor lineup includes Tencent and Xiaomi), with media offering a two-sided reading of “single major customer dependence” versus “ecosystem lock-in.”

Interpretation: With Enflame’s IPO entering the “allotment number settlement phase,” public attention shifts from wealth-creation effects to two substantive questions: the share of the raise allocated to software and ecosystem building (directly tied to the continuity of its investment in the enflame backend on the FlagOS side, where build-infra just completed its tops1.10.6 delivery loop this window) and customer/shareholder concentration. Listing pricing and first-day performance are expected to land in early September, at which point the capital narrative of the “four little dragons of domestic GPUs” will enter a new phase.

3.2 Moore Threads: MUSA Ecosystem Developers Surpass 800,000, Claims Full-Chain Embodied Intelligence Training and Deployment (09-03)

Sources: The Beijing News (Zhou Yuan: over 800,000 MUSA developers), The Paper (supply chain and embodied intelligence chain), Sohu (full embodied intelligence chain), Sohu (top-tier customers onboarded in H1)

  • On 09-03 Moore Threads gave intensive media interviews (The Beijing News/The Paper/Sohu, etc., likely tied to its first post-listing earnings communication): the number of MUSA ecosystem developers has surpassed 800,000 (per co-founder Zhou Yuan); the company has connected the full chain of embodied intelligence training and deployment, and its overall supply chain and capacity can match business development needs; in H1 it onboarded customers in key sectors including leading internet companies and telecom operators.
  • Stock market: on 09-03 Moore Threads (688795.SH) saw net main-force selling of RMB 17.1219 million (Securities Star flash report; single-day capital fluctuation of limited reference value).

Interpretation: Moore Threads’ official statements — “800,000 MUSA developers + full-chain embodied intelligence training and deployment” — form a same-day dual-track echo with GitHub-side activity in this window: on the flagos-ai side, 19 KernelGen/MThreads specialized operators were merged in half a day on 09-03 (1.3), and FlagTree fixed the mthreads3.6 TLE (1.4) — direct evidence of Moore Threads ramping up investment within the unified FlagOS software stack. The pace of completing the embodied intelligence training chain (including training operators such as amp_foreach and adaptive_max_pool3d_backward) shows its “embodied” narrative has landed at the operator level.

3.3 Hygon: Wins China Patent Gold Award; xKylin Completes Adaptation for 12 International Industrial Software Products (09-03)

Sources: Tianjin Daily (China Patent Gold Award), Drivers.com (xKylin 12 adaptations), Sina Finance (same story, reprinted)

  • On 09-03 Hygon Information (688041.SH) won the China Patent Gold Award (per Tianjin Daily) — a high-level patent award at the level of the China National Intellectual Property Administration, with Hygon recognized for CPU/DCU-related patents.
  • On 09-03, Drivers.com/Sina Finance: Hygon CPU + Kylin OS completed adaptation for 12 international industrial software products, covering CAD/CAE/AI simulation and other categories — “domestic CPUs running international industrial software” becomes a showcase for the deepening of the domestic IT innovation push.

Interpretation: Among FlagOS member units, Hygon (Hygon DCU) is a “CPU+DCU dual-track” vendor (FlagGems’ HCU backend just underwent operator performance optimization on 09-02, and FlagDNN added a hygon CPU reference implementation on 09-02). The patent gold award and industrial software adaptation news point respectively to its technical reserves and domestic-ecosystem deployment capability; these are institutional/ecosystem-side developments with no direct technical link to FlagOS but reflecting member fundamentals.

3.4 MetaX: Non-GAAP Losses for Four and a Half Straight Years, Lockup Expiry Test Looms; Appears at Space Computing Forum (09-03)

Sources: China Economic Net (non-GAAP losses), Sina Finance (lockup expiry test), Sohu (in-depth read of first semiannual report), China.com (space computing forum)

  • On 09-03, in-depth reads of MetaX’s (688802.SH) semiannual report rolled out en masse: non-GAAP net profit has been negative for four and a half straight years, with H1 operating cash flow net at -RMB 1.3 billion (China Economic Net); revenue grew nearly 45% year-on-year but profitability remains elusive; the approaching September lockup expiry of restricted shares was characterized by media as a “lockup expiry test” (expiry of lockups for initial pre-IPO shareholders).
  • At 11:01 on 09-03, China.com: MetaX appeared at the first Space Computing Industry Development Forum, themed “computing power into orbit, ground-space resonance” — space computing (satellite-borne AI compute) becomes a new narrative scenario for domestic GPUs.

Interpretation: After MetaX received its first coverage from Yi Rating yesterday (09-02), today’s media in-depth reads turned to two hard metrics — cash flow and lockup expiry: -RMB 1.3 billion operating cash flow and expected lockup-expiry selling pressure are standard tests for newly listed shares. Worth noting for comparison is its continued investment on the FlagOS side (the metax backend fixed a guard yesterday, and today both the vllm and sglang images for metax-maca3.8.1.3 were registered), with software ecosystem investment and financial pressure running in parallel — a common state among domestic GPU vendors today. Space computing is an incremental new scenario (not directly related to MetaX’s FlagOS ecosystem moves around 08-30, but the constraints that “satellite-borne” scenarios impose on operator libraries/compilers merit long-term observation).

3.5 Iluvatar CoreX and Industry Perspective: HK Shares Up Over 8%, JPMorgan Names It Top Pick; Cash Flow Pressure in “Four Little Dragons” Interim Reports Becomes Consensus (09-03)

Sources: Guandian (JPMorgan top pick), Sohu (crossing a key threshold), 21st Century Business Herald (interim report perspective on the four little dragons), Sohu (revenue surges, cash flow deteriorates), East Money (HK chip stocks rebound)

  • Iluvatar CoreX market moves: On 09-03, HK chip stocks rebounded in the afternoon, with Iluvatar CoreX (H shares) up over 8% on the day (multiple financial media outlets); a JPMorgan research note said China’s AI compute demand is growing at roughly 80% CAGR, naming Iluvatar CoreX, SMIC and others as top picks — sell-side analysts giving high-growth estimates for long-term demand in China’s AI compute sector.
  • Industry perspective (two pieces around midday on 09-03): Sohu’s “Domestic AI Chip Revenue Surged in H1 — Why Is Operating Cash Flow Deteriorating at Most Vendors?” and 21st Century Business Herald’s “Interim Report Perspective on the ‘Four Little Dragons’ of Domestic AI Chips: Two Turn Profitable on Paper, Cash Flow Still Under Pressure” — the common conclusion across the “four little dragons” (Enflame/MetaX/Moore Threads/Iluvatar CoreX, etc.) interim reports is the coexistence of high revenue growth on paper and operating cash flow pressure, in tune with the discussion in 3.1/3.4; Sohu’s “From Technological Breakthrough to Scale Revenue, Iluvatar CoreX Crosses a Key Threshold” (early on 09-03) offers a positive reading of a “scale revenue inflection point.”

Interpretation: Iluvatar CoreX (a FlagOS member, Iluvatar backend) is getting positive pricing in capital markets — “JPMorgan top pick + single-day gain over 8%” — alongside the industry’s cautious “cash flow pressure” narrative: the domestic AI chip sector has entered a phase of “tug-of-war between the demand story (80% CAGR) and financial reality (cash flow).” For FlagOS, sell-side high-demand estimates mean the market-space narrative for a unified software stack (adapt once, deliver across multiple chips) continues to strengthen; a rebound in sector sentiment also helps member units maintain the intensity of their investment in the open-source software stack.

IV. Summary

In this window (09-03 10:18 ~ 09-04 10:18), the main threads of the FlagOS ecosystem:

  1. High GitHub-side activity continues, release quiet period unchanged: 14 repos pushed, 51 in-window commits; tags/releases verification shows no new version actions, and the engineering cadence continues after the 2.2 rc0 validation period wrapped up. The biggest highlight is build-infra completing its first full member delivery loop of “rollback-reregister-mark delivery”: after the Enflame vLLM 0.24.0 delivery record was revoked, it was verified with dual versions tops1.9.10/tops1.10.6, and finally completed F+T validation delivery with 1.10.6 (#703→#715/#716→#717).
  2. Ascend toolchain switches wholesale to CANN 9.0.0: build-infra registered the Ascend sglang application image with cann9.0.0 (#713), and FlagGems CI removed cann850 the same day (#5970) — the upgrade from 8.5.0 to 9.0.0 lands simultaneously on both the operator library and image management sides.
  3. Moore Threads adaptation enters the “dense expansion of specialized operators” phase: FlagGems-Experimental merged 19 KernelGen/MThreads specialized operators in half a day (#291~#309), and FlagTree fixed the mthreads3.6 TLE in parallel; on the same day Moore Threads officially announced over 800,000 MUSA ecosystem developers and full-chain embodied intelligence training and deployment — GitHub commits and official narrative corroborate each other.
  4. DeepSeek inference pipeline continues to be optimized: on the FlagGems-vllm KMCompiler side, DeepSeek-v3.2 grouped_topk performance optimization was completed, the Fused MoE general operator was aligned with FlagGems upstream, and platform work converged onto vendor backends; the sglang/vllm image matrix added MetaX metax and Ascend registrations, further expanding multi-vendor inference image coverage.
  5. Member financial news concentrated on the Enflame IPO: allotment numbers announced (20,657, 500 shares per lot), with the RMB 6.1B raise allocation and “Tencent dependence” in focus; continued interim report reads and market pricing for Moore Threads/MetaX/Iluvatar CoreX (JPMorgan top pick on Iluvatar CoreX, MetaX lockup expiry approaching) form the sector panorama — high revenue growth coexisting with cash flow pressure.
  6. News side quiet for a third consecutive window: zero direct hits at the component level; on the BAAI institutional side there are routine developments such as the first cohort graduation of the “AI+ Applications and Governance Pilot Program”; no new official releases to carry forward the ecosystem diffusion of the Day0 article.

Appendix: Complete Source List

Source Verification Result
GitHub org repos API (52 repos) 14 repos pushed within the window, no new repos created; the pushed_at change for docs/release-info/FlagTree-AscendNPU-IR was verified via commits as non-default-branch activity (0 commits on default branch), excluded from the main text
GitHub commit search All 51 commits within the window retrieved (FlagGems-Experimental 19, build-infra 17, FlagGems 4, FlagTree 3, TransformerEngine-FL 3, FlagGems-vllm 3, sglang-plugin-FL 2, FlagCX/Torch-FL/FlagDNN 1 each); vllm-plugin-FL/docs default branch had 0 commits within the window
GitHub tags/releases API No new tags, no new releases within the window: latest releases of active repos were all created before the window (FlagGems v5.3.0 06-24, FlagCX v0.13.0 06-24, vllm-plugin-FL v0.3.0-rc0 08-24, sglang-plugin-FL v0.1.0 06-24, FlagGems-vllm v0.1.1-rc0 08-11, etc.); the 2.2 release quiet period continues
build-infra #699-#717 series Verified item by item: Enflame delivery revocation (#703) → tops1.9.10/tops1.10.6 re-registration (#715/#716) → F+T marked delivery (#717); removal of distrust flagtree wording (#702); metax vllm/sglang image records (#700/#712); ascend-cann9.0.0 sglang record (#713); sglang runtime fixes (#699/#704/#706/#707/#710); record_only (#711); plugin commit derived+asserted (#709); semantic anchor gating (#701)
FlagGems-Experimental #291-#309 19 [KernelGen][MThreads] specialized operator items verified one by one (merged in a concentrated burst from 12:51 to 16:48)
FlagGems #4128/#5948/#5963/#5970 SiliconFlow cumprod precision and inplace integer fix, setup.sh image installation, use_gems CI expansion, removal of cann850
FlagTree #1092/#1094/#1095 AMD/ROCm volatile retention, mthreads3.6 TLE tle_supported fix, FlagOS copyright header update
FlagGems-vllm #693/#92 and 3 others Hopper w8a8 block fp8 matmul, Fused MoE upstream alignment + vendor backend consolidation, DeepSeek-v3.2 grouped_topk optimization
TransformerEngine-FL #108-#110, FlagCX #566, sglang-plugin-FL #43/#83, Torch-FL #258, FlagDNN FlagOS/Reference Adam interface completion and compute-scale alignment, UIL txda_stream support, empty device (xinchuang) support, uint type conversion, cpu_op tests
Google News RSS Chinese group FlagOS/component-level zero hits in 24h (third consecutive window); “BAAI/BAAI” query hits were all BAAI community reposts or iPolloWork OPC coverage (verified that iPolloWork is an independent Devin-AXIS Agent project, unrelated to BAAI/FlagOS, excluded); multiple items verified for Enflame IPO (20657 winning lots, RMB 6.1B fundraising use, Tencent dependency response, RMB 61.1B issue market cap), Moore Threads (MUSA 800K developers, embodied intelligence pipeline, supply chain), Hygon (patent gold award, xKylin adaptation for 12 industrial software titles), MetaX (recurring non-GAAP losses, cash flow -RMB 1.3B, lock-up expiry, space computing forum), Iluvatar CoreX and industry (Hong Kong stocks +8%, JPMorgan top pick, interim report insights on the four little dragons)
Google News RSS English group FlagOS/BAAI when:2d-3d all empty; component queries hit only pre-window old news, no valid new hits
HN Algolia 1 hit each for FlagOS/FlagGems, both noise (YC match safety, Zenodo link skepticism), excluded
BAAI community (gnews hits) “AI+ Application and Governance Leadership Program” first cohort successfully completed (09-03 23:30); the rest are third-party content reposts (Gemini 3.8/Yan Shuicheng/embodied intelligence, etc.), not BAAI’s own achievements
The Beijing News/The Paper/Sohu/China Economic Net/Sina Finance/China.com/East Money/Securities Times/Cailian Press/EET China/21st Century Business Herald/Guandian/Drivers Home/Tianjin Daily Member unit entries verified one by one by title + summary; “500 shares subscribed per winning lot number” and “20,657 winning lot numbers” follow the Securities Times/East Money wording