Research window: 2026-08-31 10:18 ~ 2026-09-01 10:18 Beijing Time Sources: GitHub (org: flagos-ai, 52 repos + 76 commit search results + releases/PR verification), Google News RSS (multiple Chinese and English queries, proxy chain normal), HN Algolia, Tavily/web cross-verification (see appendix for details)

Index

  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 Headline: FlagOS 2.2 RC0 release checklist finalized, 24 components lock version lines by tier (09-01)
    • 1.2 FlagSparse releases v0.3.0-rc0.post1, the first RC verification tag for 2.2 (09-01)
    • 1.3 build-infra: megatron training validation expands to Enflame/Iluvatar dual backends, RL framework shifts to verl (08-31/09-01)
    • 1.4 Torch-FL: HuggingFace official tests pass accelerator-by-accelerator per model, model coverage methodology takes shape (08-31)
    • 1.5 FlagTree / FlagGems: multi-backend operators and CI advance on two fronts (08-31)
    • 1.6 Other repos: plugin-layer CI migration, CICD image updates, org docs restructured by SIG (08-31/09-01)
  • II. News Coverage and Ecosystem
    • 2.1 BAAI scholars embodied intelligence workshop: embodied intelligence has not reached its “GPT moment,” bottleneck lies in physical-space understanding and standards (reported 08-31)
    • 2.2 2026 BAAI Conference approaching: embodied intelligence and humanoid robotics forum agenda published (08-31)
    • 2.3 Industry backdrop: high-end compute orders booked out 3 years, domestic AI chip semi-annual reports collectively strong (08-31)
  • III. Member Company Deep Dive
    • 3.1 Enflame: IPO issue price set at 142.18 yuan/share, strategic placement Tencent adds 248 million, subscription on September 2 (08-31/09-01)
    • 3.2 Hygon: wins China Patent Gold Award, completes joint adaptation and tuning with Tencent Cloud Cube Sandbox (08-31)
    • 3.3 MetaX: follow-up on semi-annual report — still loss-making on non-recurring basis, receivables and cash flow under pressure (08-31)
    • 3.4 Iluvatar: interpretation of turnaround via investment income continues to gain traction, revenue up 191.6% (08-31)
    • 3.5 Tsingmicro: tsingmicro vLLM 0.24.0 validation formally recorded, application image enters rc line (08-31/09-01)
  • IV. Summary
  • Appendix: Complete Source List

1. Open-Source Project Progress (GitHub Activity)

Window Overview: Of the 52 repos in the org, 28 had pushes during the window (including multiple PR branches and rc branches), and commit search returned 76 in-window commits. Main threads this window: 1) The FlagOS 2.2 RC0 release manifest officially landed (community #103 merged into main) — 24 components locked to rc0 validation versions across four tiers L0-L3, the first full-stack version freeze since 2.1, while also surfacing an alignment issue over the release scope of verl-related components; 2) FlagSparse released v0.3.0-rc0.post1, becoming the first publicly released RC validation tag in the 2.2 cycle; 3) build-infra megatron training validation expanded — Enflame (enflame-tops1.10.6) and Iluvatar (iluvatar-corex4.5.0) dual backends passed megatron_training E2E under both compiler paths, while megatron_rl was marked paused and the RL framework shifted as a whole to verl, echoing the verl issue in the RC0 manifest; 4) The Torch-FL model coverage validation methodology took shape — HuggingFace official tests run per-model through the accelerator.

1.1 Headline: FlagOS 2.2 RC0 release manifest lands, 24 components lock version lines by tier (09-01)

Sources: community #103 (merge commit 3679f77), release-2.2-rc0.yaml

  • 09-01 01:29 (Beijing time) community #103 merged into main: added release/2.2/release-2.2-rc0.yaml, listing all 24 source entries for FlagOS 2.2 RC0 in vcstool format, along with the manage-release.py workflow (release-branch-tag.yml). The 2.2 cadence continues 2.1’s “per-module version numbers + rc0.postN validation iterations” model: each successful validation tags an incrementing tag and updates the manifest version.
  • L0 infrastructure tier: flagtree split into three lines by Triton version (triton3.6/3.5/3.3, all 0.7.0rc0.post1+tritonX, baseline current version 0.6.0+tritonX; triton3.6 shares commit 2e62581 with main, i.e. yesterday’s FlagTree #1043); flagcx v0.14.0-rc0.post1 (current v0.13.0).
  • L1 compute library tier: flaggems v5.4.0-rc0.post1 (current v5.3.5, 2.1 GA was v5.3.0), flagfft v0.2.0-rc0.post1 (current v0.1.0), flagsparse v0.3.0-rc0.post1 (current v0.2.0, consistent with the same-day release, see 1.2), flagdnn/flagblas/flagtensor/flagaudio all v0.3.0-rc0.post1, flagattention v0.4.0-rc0.post1, flaggems-vllm v0.2.0-rc0.post1, flaggems-sglang v0.1.0-rc0.post1 — a new 2.2 entry (the repo previously had no tags; its first version starts at 0.1.0).
  • L2 framework adaptation tier: torch-fl v0.2.0-rc0.post1 (the manifest notes it was named pytorch-plugin-fl in 2.1, and the repo has been renamed PyTorch-Plugin-FL → Torch-FL); vllm-plugin-fl v0.3.0-rc0.post1 (current v0.2.1); sglang-plugin-fl v0.2.0-rc0.post1; transformerengine-fl/megatron-lm-fl v0.3.0-rc0.post1; flagos-compressor v0.1.0-rc0.post1 — a new 2.2 module (FEP #91 quantization framework).
  • L3 application/tooling tier: flagscale v2.1.0-rc0.post1 (current v2.0.0), kernelgen v2.2.0-rc0.post1 (current v2.1.0), kernelgenbench v0.2.0-rc0.post1, flagrelease v0.3.0-rc0.post1.
  • Open review issue (P1): the manifest does not include verl-fl / verl-hardware-plugin, yet FEP 0062 (verl-hardware-plugin v0.1.0 multi-chip GRPO E2E) remains in Implementable status with Target Version FlagOS 2.2, and tracking issue #73 is also in the 2.2 milestone — the misalignment between release scope and the FEP/milestone/roadmap was explicitly raised in review, and it must be confirmed whether the 2.2 deliverables will ultimately include the verl family of components.

Interpretation: The 2.2 RC0 manifest is the first full-stack version freeze since 2.1 (six components released in sync on 2026-06-24), with three highlights: first, two new modules — flaggems-sglang and flagos-compressor (quantization framework) — enter the official release manifest for the first time; second, the PyTorch-Plugin-FL → Torch-FL rename is formally confirmed in the manifest; third, FlagTree’s split management by Triton version line (3.6/3.5/3.3) becomes the release baseline. The verl alignment issue and the build-infra side’s “megatron_rl paused, RL shifts to verl” (see 1.3) corroborate each other, indicating that the 2.2 RL framework roadmap is re-converging and the manifest may still change before the official release.

1.2 FlagSparse releases v0.3.0-rc0.post1, the first RC validation tag of 2.2 (09-01)

Source: FlagSparse Release v0.3.0-rc0.post1

  • 09-01 02:02 (Beijing time, i.e. 08-31 18:02 UTC) FlagSparse released v0.3.0-rc0.post1 (the only release in the org this window), corresponding to the flagsparse entry in the 2.2 RC0 manifest (current version v0.2.0 → rc0 target 0.3.0), and the first publicly released RC validation tag in the 2.2 cycle.
  • The 0.3.0-rc0 branch saw continuous commits during the window (including NCIC-AlphaSparse-related contributions), with validation iterations proceeding on an “rc0.post1 → rc0.post2” cadence.

Interpretation: FlagSparse (sparse operator library) is the first to complete the initial round of 2.2 RC0 validation and tag a release, the first instance of the manifest workflow (validation passes → tag → update manifest) running end to end; as sparse computation’s share in MoE large models (such as Hy4-class Gated MLA structures) keeps rising, FlagSparse’s release cadence is worth watching.

1.3 build-infra: megatron training validation expands to Enflame/Iluvatar dual backends, RL framework shifts to verl (08-31/09-01)

Sources: build-infra #634, #635, #636, #637, #640, #641, #642, #643, #646, #647, #648, #654, #655, #657, #658

  • 08-31 14:24 #634 merged: megatron 0.17.1 training validation expanded — enflame-tops1.10.6 and iluvatar-corex4.5.0 dual backends passed megatron_training mock-data pretrain_gpt.py 5-iter E2E, and passed under both the flagtree + triton compiler paths, with consistent F/T validation loss (enflame 1.088891E+01, iluvatar 1.087054E+01). The Enflame side required two in-container patches (flagtree jit_fuser noop, ECCL rejecting fp64 all_reduce — Enflame-specific, as other vendors’ collectives accept torch.double), not persisted into the wheel; Iluvatar corex4.5.0 is a newly added matrix row (coexisting with the existing corex4.4.0), passing cleanly with zero workarounds.
  • 08-31 14:31 #635: megatron_rl marked paused — the status matrix explicitly states the RL framework is shifting to verl (the Verl-FL line), and megatron_rl will no longer be advanced; 08-31 23:22 #654 added an upstream PR tracking matrix for the enflame/iluvatar/metax training backends (bidirectionally linking upstream workarounds and FlagOS-side records).
  • 08-31 17:57 #640: Enflame vendor defect report — two Enflame defects found during megatron 0.17.1 validation were turned into a handoff document: torch-gcu all_gather_object returning garbage values (TOPS 1.9.10), and ECCL rejecting fp64 all_reduce (TOPS 1.10.6), with links to upstream MLF workaround PRs.
  • 08-31 19:50-22:50 megatron wheel/image build chain fixes: #641 changed the default pypi_repo to per-vendor flagos-pypi-{vendor}, #642/651 corrected the –no-cache semantics for application images, #650/652/657 registered multiple application image tags (enflame-tops1.9.10/1.10.6, iluvatar-corex4.4.0/4.5.0).
  • 08-31 20:52-21:51 Tsingmicro vLLM 0.24.0 validation formally recorded (#643 docs record, #646 registered tsingmicro-tsm260610 deps_app, #648 pinned PR #421 head wheel); 09-01 00:29 also registered the vLLM 0.3.0rc0 application image tag for tsingmicro-tsm260610 — the Tsingmicro inference stack advances toward the rc line (see 3.5).
  • 08-31 23:27 #655: parameterized the vLLM serve-test time budget (adapting to slow cold starts); 09-01 08:24 #658: added a backend notes column to the status matrix infrastructure table.
  • 08-31 11:56 #633: eliminated duplicate PATH entries in build images (runtime fix).

Interpretation: The main thread for build-infra this window is megatron training validation moving from “single backend working” to a “multi-backend dual-compiler matrix” — Enflame tops1.10.6 and Iluvatar corex4.5.0 completed F+T dual-path validation on the same day, with Iluvatar requiring zero workarounds; Enflame’s two container-level defects were formally documented as a vendor handoff (with links to upstream fix PRs), marking the validation process’s entry into a mature loop of “find defect → report to vendor → track upstream.” The pausing of megatron_rl and the shift to verl, together with the verl alignment issue in the RC0 manifest, jointly point to VeRL-FL as the primary RL training framework in 2.2.

1.4 Torch-FL: HuggingFace official tests run per-model through the accelerator, model coverage methodology takes shape (08-31)

Sources: Torch-FL #236, #230, #231, #222, #234

  • 08-31 22:39 #236 merged: HuggingFace official per-model test runner — downloads and caches the Transformers source version exactly matching the installed wheel, injects the TRANSFORMERS_TEST_DEVICE_SPEC custom device spec, runs the full official test suite under tests/models/<architecture>/ for each model architecture in a separate pytest subprocess, and outputs a structured JSON report (per-test status + overall verdict). Currently report-only mode; baseline promotion, fingerprint deduplication, operator attribution, and automatic issue creation are listed as follow-up work; the runner is backend-neutral (default device flagos, --device reusable for other PrivateUse1 accelerators). The PR was drafted with assistance from Claude Code CLI (Claude Opus 5) and human-reviewed.
  • 08-31 12:15 #230 merged: test: probe transformers models per architecture on accelerators — per-architecture model probing, forming a two-layer coverage scheme with #236 (“probing + official tests”).
  • 08-31 13:12 #231 merged: MUSA backend empty-operand fill_/zero_/cat take the device path (Moore Threads); #222 merged: Ascend profiler device-time path accepts fused matmul (Ascend).
  • 08-31 14:14 #234 merged: rerouted stale FlagGems Python qualnames.

Interpretation: Yesterday (#229) built skill docs for model coverage research, and today the “official test runner” landed — Torch-FL is advancing extremely fast on the main validation thread of “running official PyTorch/Transformers tests on accelerators.” Once the mechanism of per-architecture + official test suite + structured reports is working, each chip’s “real model compatibility” will for the first time have a quantifiable, reproducible metric — key infrastructure for comparative evaluation of domestic accelerators, and the engineering support for the 2.2 model coverage deliverables.

1.5 FlagTree / FlagGems: multi-backend operators and CI advance on two fronts (08-31)

Sources: FlagTree #1068, #1070, #1043, FlagGems commits

  • FlagTree: 08-31 15:17 #1068 synced the MTHREADS backend to musa commit 43436c4c (Moore Threads); 18:04 #1070 added GitHub Actions CI for FlagGems tests (cross-repo test infrastructure connected); 20:30 #1043 Nvidia TritonGPU support for interleaved concat, heterogeneous segmentation, and segment merging — this commit (2e62581) is also designated by the 2.2 RC0 manifest as the baseline for the flagtree triton3.6 line.
  • FlagGems (14 commits in a single day, the most active repo in the org): 4 on the KMCompiler side (iluvatar sparse_sampled_addmm K<16 compile fix #5705, igammac operator registration #5831, out-of-place exponential registration + benchmark baseline #5843, linalg_solve_triangular optimization #5753); 3 on the MTHREADS side (matmuladd reusing addmm SQMMA/FMA kernels #5781, linear performance optimization #5863, dispatch_key corrected to PrivateUse1 #5865); on the Nvidia side Hopper MM dispatch and tuning #5780, KernelGen added an nll_loss2d Triton kernel #5660; on the infrastructure side a pure-triton normcdfinv fallback (non-CUDA backends) #5864, tool quick mode #5866, rule-check results reported to a Feishu Bitable #5854.

Interpretation: FlagTree and FlagGems maintain a dual cadence of “multi-backend operators + CI infrastructure”; FlagGems’ rule-check Feishu reporting shows the quality gate has been wired into the team collaboration flow; FlagTree #1070’s cross-repo CI (FlagGems tests running on FlagTree) is an infrastructure signal of coordinated validation between the compiler stack and the operator library.

1.6 Other repos: plugin-layer CI migration, CICD image updates, org docs restructured by SIG (08-31/09-01)

Sources: vllm-plugin-FL #415, FlagGems-vllm #725, sglang-plugin-FL #79, flir #70, Megatron-LM-FL #129, .github docs

  • vllm-plugin-FL 08-31 17:26 #415: CI-related code migrated from release/0.2 to main (branch convergence).
  • FlagGems-vllm 08-31 14:09 #725: fixed grid.x overflow in the enflame backend scaled_int8_quant static path.
  • sglang-plugin-FL 08-31 18:58 #79: bypassed the torch_musa isin operator and fixed VL precision (Moore Threads).
  • flir 08-31 18:23 #70: fixed UseAnalysis cloning MixUse AllocOp in the TLE path.
  • Megatron-LM-FL 09-01 10:06 (Beijing time) #129: CICD platform test image update.
  • .github 08-31 10:58-13:16: Core Components docs restructured by SIG (added FlagGems-vllm/FlagGems-sglang under sig-operator, component table converted to HTML), org governance docs refreshed in sync with the 2.2 component landscape.
  • community 08-31 12:02-13:48 #102: sig-edge activated — SIG Edge charter and design docs landed (see the ecosystem-side connection in 2.2).

II. News Coverage and Ecosystem

2.1 BAAI Scholars Embodied Intelligence Symposium: Embodied Intelligence Has Not Yet Reached Its “GPT Moment”; the Bottleneck Lies in Physical-Space Understanding and Standards (Reported 08-31)

Source: MyDrivers (gnews repost)

  • On 8/27, the “BAAI Scholars · Embodied Intelligence Academic Symposium” was held at Longfusi in Beijing (widely reposted by financial media on 8/31). More than 20 BAAI scholars from Peking University, Tsinghua University, Beihang University, Beijing Jiaotong University, and the Institute of Automation of the Chinese Academy of Sciences held discussions across three themed roundtables: “World Models, Embodied Intelligence, and Dexterous Manipulation.”
  • Core judgment: Embodied intelligence has not yet reached its “GPT moment.” The key constraint is not the robot hardware itself, but the structured understanding and reasoning capability in physical space, as well as the high migration costs caused by the lack of unified standards among the robot body, models, and data. Several panelists favored a layered architecture, exploring modular collaboration among large models, world models, and VLA. In terms of deployment pace, vertical scenarios such as industrialization are expected to commercialize first, while general intelligence deployment will require two to three years or even longer.

Interpretation: As the lead organization behind FlagOS, BAAI’s academic layout in embodied intelligence directly echoes the direction of FlagOS-Robo (the embodied intelligence toolchain). The consensus on the “lack of unified standards” also explains the methodological value of FlagOS in converging multi-chip/multi-body support through a unified software stack.

2.2 2026 BAAI Conference Approaching: Agendas for the Embodied Intelligence and Humanoid Robotics Forums Released (08-31)

Source: BAAI Hub 2026 BAAI Conference Agenda, Embodied Intelligence CEO Forum Agenda

  • The 2026 BAAI Conference is in countdown (around mid-September), and the agenda for the Embodied Intelligence and Humanoid Robotics Forum has been released: Wang He of Galbot serves as forum chair, and Wang Pengwei, head of BAAI’s Embodied Model Research Center, will present the world foundation model ORCA, embodied brain RoboBrain, and cross-body brain-cerebellum coordination framework RoboOS. A separate “Embodied Intelligence CEO Summit” roundtable will also be held (Zhou Yong of Lingxin Qiaoshou, Liu Dong of Xingyuanzhi, Han Fengtao of Qianxun Intelligence, and others).
  • The BAAI Hub platform maintained high-frequency updates during this window (paper interpretations on machine learning protein engineering, automated transition-state generation, etc.), remaining active on the content side as the lead organization behind FlagOS.

Interpretation: BAAI’s embodied intelligence portfolio (ORCA/RoboBrain/RoboOS) and the FlagOS-Robo toolchain both belong to the BAAI ecosystem. If the conference announces new embodied/software stack developments, it will be a key item to watch in subsequent windows.

2.3 Industry Background: High-End Compute Orders Booked Out Three Years Ahead; Domestic AI Chip Interim Reports Uniformly Strong (08-31)

Source: Sina Finance (domestic compute orders), Cailian Press (AI chip interim reports)

  • High-end compute remains in persistent shortage: multiple domestic compute companies have orders booked out three years ahead, Biren Technology’s revenue grew nearly 20-fold, and Cambricon/Moore Threads saw sharp revenue increases. “Domestic AI chip interim reports uniformly strong; compute shortage drives industry boom” became the main theme in financial media on 8/31 (reports were密集 on comparisons of the interim reports of the domestic GPU “four little dragons,” MetaX turning profitable, Enflame’s IPO, etc.; items related to member units have been included in Chapter III).
  • Component-level queries for FlagOS (FlagOS/FlagGems/FlagScale/FlagTree/FlagPerf/FlagAttention/FlagCX/KernelGen, etc., in both Chinese and English) recorded zero hits for the fifth consecutive day in this window. Direct news coverage of FlagOS remains in a quiet period; GitHub commits and member-unit financial developments are currently reliable sources of updates.

III. Member Unit Deep Dive

3.1 Enflame: IPO price set at RMB 142.18/share, strategic placement Tencent adds RMB 248M, subscription on September 2 (08-31/09-01)

Sources: 21st Century Business Herald/China Securities Journal (strategic placement revealed), Sina Finance (issuance announcement), DoNews

  • On the evening of 8/31, Enflame Technology disclosed its STAR Market IPO issuance announcement: issuance price RMB 142.18/share, 43,035,173 shares issued, expected to raise RMB 6.119 billion, market cap at listing approximately RMB 61.187 billion (price-to-sales valuation chosen, PS ratio of 61.8x below peers); online and offline subscription date is September 2 (RMB 71,100 payment per winning lot), with an online roadshow held on 9/1.
  • Strategic placement results: Tencent’s Shanghai Qishan Investment subscribed 1,747,100 shares (4.06% of issuance, corresponding to RMB 248 million, with a 36-month lock-up alongside the employee asset management plan) — top customer Tencent further increasing its stake; Xiaomi (Wuhan 1810) and GigaDevice (Shenzhen Waitan Technology) each received 492,300 shares (1.14%); Tongfu Microelectronics and ZTE each received 562,700 shares (1.31%); Shanghai Huahong (Xinzhi Times) and Shanghai Yidian (Aidixi) each received 514,300 shares (1.2%); the National Social Security Fund and Basic Pension Insurance Fund participated; sponsor subsidiary Zhongzheng Investment co-invested 860,700 shares (2%, 24-month lock-up).
  • Offline inquiry heat: 337 offline investors with 11,805 allocation targets participated, with quotes ranging from RMB 14.36 to 159.32, and an overall offline subscription multiple of 2,648.93x.
  • Prospectus highlights: hardware based on proprietary instruction set, GCU-CARE accelerated computing unit and GCU-LARE inter-chip interconnect benchmarked against Tensor Core/NVLink; software side features self-developed full-stack “TopsRider” (driver, compilation language and compiler, operator library, toolchain), not following the CUDA ecosystem — echoing the GCU stack within FlagOS (FlagCX collective, FlagTree gcu backend, torch.compile).

Analysis: Yesterday’s report noted “IPO registered, 9/2 subscription”; this window completes the trifecta of issuance pricing, strategic placement finalization, and roadshow, putting Enflame just one step from listing; the strategic placement roster brings together Tencent/Xiaomi/GigaDevice/ZTE/Tongfu/Huahong/Yidian/Social Security, with clear intent to bind the industry chain (locking in wafer, memory, and packaging capacity + full-stack computing equipment synergy). As a FlagOS member unit, the synergy between its “TopsRider” and the FlagOS open-source stack is a point to watch going forward.

3.2 Hygon: Wins China Patent Gold Award, completes joint adaptation and tuning with Tencent Cloud Cube Sandbox (08-31)

Sources: Sina Finance (Patent Gold Award), Sohu (Cube Sandbox joint adaptation)

  • 8/31 Hygon Information won the 26th China Patent Gold Award (domestic server security direction), another landmark recognition of “the hard bone of domestic server security.”
  • 8/31 Hygon completed deep joint adaptation and tuning with Tencent Cloud open-source project Cube Sandbox — another win for DCU-side cloud-native/sandbox environment adaptation (continuing the pace of yesterday’s Hy4 preview Day 0 adaptation).
  • Capital market side: Hygon became the stock with the largest net buying by public funds in the first half of the year (8/31 report).

Analysis: Hygon is heating up simultaneously across three lines — “model Day 0 adaptation (Hy4) → open-source project joint adaptation (Cube Sandbox) → patent and capital recognition”; its DCU official PyTorch core route within FlagOS (Torch-FL #227) continues to benefit from this high level of activity.

3.3 MetaX: Follow-up interpretation of semi-annual report — still loss-making on a non-recurring basis, receivables and cash flow under pressure (08-31)

Sources: Zhiding (operating cash flow), Sohu (semi-annual report breakdown), Securities Star (Jiangsu Guoxin)

  • Follow-up breakdowns of MetaX’s semi-annual report (yesterday reported turnaround: net profit RMB 612 million, revenue RMB 1.324 billion +44.67%) appeared densely: Zhiding noted net profit of RMB 612 million but operating cash flow still outflowing nearly RMB 1.3 billion; Sohu’s breakdown showed a loss of RMB 48.86 million on a non-recurring basis, accounts receivable near RMB 1 billion, market cap RMB 270 billion; “domestic GPU big three/four little dragons semi-annual report comparison” articles listed it as a sample.
  • Jiangsu Guoxin (8/31) responded: the company has not directly or indirectly invested in MetaX chips (clarifying market rumors).

Analysis: Beneath MetaX’s “turnaround to profitability” halo, non-recurring losses, operating cash outflow, and high receivables have become the focus of market scrutiny — the quality of the core business during the domestic GPU commercialization realization period still needs time to verify; its MetaX backend adaptation within FlagOS (MUSA line) maintains steady progress (see 1.4/1.5).

3.4 Iluvatar CoreX: Interpretation of turnaround via investment gains continues to ferment, revenue up 191.6% (08-31)

Sources: JRJ (turnaround via stock trading gains), East Money (revenue +191.6%), Sohu (from RMB 600M loss to RMB 100M profit)

  • On 8/31, media continued to break down Iluvatar CoreX’s semi-annual report: revenue +191.6% but the turnaround mainly relied on over RMB 700 million in equity investment gains (JRJ headline directly pointed to “turnaround via stock trading gains, core business still not profitable”); the quality of the “turnaround” from a RMB 600 million loss to RMB 100 million profit was questioned from multiple angles.
  • In contrast to the financial side: Iluvatar CoreX’s corex4.5.0 backend within FlagOS completed megatron training dual-compiler path verification (zero workarounds, see 1.3) — ecosystem adaptation progress clearly ahead of financial realization.

Analysis: Iluvatar CoreX continues yesterday’s interpretation thread (turnaround dependent on investment gains), with this window adding a megatron training verification milestone; the FlagOS-side verification loop (yesterday’s vLLM wrap-up + today’s megatron) resonates with its core business ramp-up, with ecosystem value realized ahead of financial value.

3.5 Tsingmicro: tsingmicro vLLM 0.24.0 verification formally recorded, application image enters rc line (08-31/09-01)

Sources: build-infra #643, #646, #648, commit 7562122

  • 08-31 20:52-21:51: Tsingmicro (tsingmicro-tsm260610) vLLM 0.24.0 verification formally recorded (#643), with deps_app registered (#646) and verification pinned to vllm-plugin-FL PR #421 head wheel (#648) — yesterday’s #631 had just introduced the tsm_smi device check into the verification script, and this window completed the first round of 0.24.0 verification closure.
  • 09-01 00:29: Registered the vLLM 0.3.0rc0 application image tag for tsingmicro-tsm260610 (2.1.2-0.3.0rc0_gbd010ce.d20260831) — Tsingmicro’s inference stack entering the rc application image line for the first time.

Analysis: From “entering the verification pipeline” (yesterday) to “0.24.0 verification record + rc image tag” (today) in just one day, Tsingmicro is the fastest-onboarding member unit chip recently; its vLLM application image is now managed on the same line as Enflame/Iluvatar CoreX, with inference stack engineering verification comprehensively accelerating.

IV. Summary

In this window (08-31 10:18 ~ 09-01 10:18), the FlagOS ecosystem main threads:

  1. FlagOS 2.2 RC0 release manifest finalized is the biggest event: 24 components locked to rc0 verification versions across L0-L3 tiers, with flaggems-sglang and flagos-compressor (quantization framework) entering the formal release manifest for the first time, and the PyTorch-Plugin-FL → Torch-FL rename confirmed in the manifest; the review also raised alignment issues between the verl-family components (verl-fl/verl-hardware-plugin) release scope and FEP 0062/milestone, leaving room for adjustment in the 2.2 manifest.
  2. FlagSparse v0.3.0-rc0.post1 released as the first external RC verification tag of the 2.2 cycle, with the manifest’s “verify → tag → update manifest” workflow running through its first instance.
  3. build-infra megatron training verification expanded: Enflame tops1.10.6 and Iluvatar CoreX corex4.5.0 dual-backend, dual-compiler paths (flagtree + triton) passed E2E on the same day, with Iluvatar CoreX achieving zero workarounds; two Enflame container-level defects formed vendor handoff documents with links to upstream fix PRs; megatron_rl marked paused, with the RL framework overall pivoting to verl.
  4. Torch-FL model coverage verification methodology took shape: HuggingFace official tests run model-by-model on accelerators (#236), with “probing + official tests” dual-layer coverage providing a quantifiable measure of real model compatibility for domestic accelerators.
  5. News-side component-level queries hit zero for the fifth consecutive day; member unit financial news was dense: Enflame IPO priced at RMB 142.18/share + strategic placement Tencent increasing stake (9/2 subscription), Hygon Patent Gold Award + Cube Sandbox joint adaptation, MetaX/Iluvatar CoreX semi-annual report quality breakdowns fermenting; Tsingmicro vLLM 0.24.0 verification record + rc image tag landed.
  6. Ecosystem signals: BAAI embodied intelligence academic layout active (scholars’ symposium + BAAI Conference agenda publicized), echoing the FlagOS-Robo direction; industry prosperity with high-end computing orders booked out 3 years provides support for member units’ fundamental upward trajectory.

Appendix: Complete Source List

Source Verification Result
GitHub org repos API (52 repos) 28 repos pushed within window, all 76 commits verified
GitHub commit search 76 commits within window (all 3 pages retrieved), no omissions
GitHub releases API Only release: FlagSparse v0.3.0-rc0.post1
community #103 / release-2.2-rc0.yaml All 24 manifest entries verified item by item (L0-L3 four tiers, versions and new modules); P1 review comments (verl alignment) recorded
build-infra #634-658 series megatron verification/defect reports/image chain/Tsingmicro vLLM records verified item by item
Torch-FL #236/#230/#231/#222/#234 Titles/status/merge times verified item by item; #236 body verified (HF official test runner mechanism)
FlagTree/FlagGems/flir/vllm-plugin-FL/FlagGems-vllm/sglang-plugin-FL/Megatron-LM-FL/.github Commit titles and PR statuses verified
Google News RSS Chinese group FlagOS/component queries zero hits; Enflame (issuance price/strategic placement/roadshow), Hygon (Patent Gold Award/Cube Sandbox), MetaX (semi-annual report breakdown), Iluvatar CoreX (turnaround interpretation), BAAI (embodied symposium) multiple items verified
Google News RSS English group No valid hits
HN Algolia Zero hits
21st Century Business Herald/China Securities Journal (Enflame strategic placement) Body verified: RMB 142.18/share, RMB 6.119 billion, Tencent RMB 248 million, strategic placement roster and lock-up periods
Sina Finance (essence of four major securities newspapers) Issuance announcement details verified (43,035,173 shares, market cap RMB 61.187 billion, TopsRider software stack)
Kuaikeji (BAAI scholars’ symposium) Body verified: 8/27 Longfusi, three roundtables, core judgments
BAAI community hub.baai.ac.cn (BAAI Conference agenda) ORCA/RoboBrain/RoboOS and forum guests verified
Zhiding/Sohu/JRJ/East Money (MetaX, Iluvatar CoreX) Key figures verified: MetaX non-recurring loss RMB 48.86 million, cash outflow RMB 1.3 billion; Iluvatar CoreX +191.6%