Research window: 2026-09-01 10:18 ~ 2026-09-02 10:18 Beijing Time Sources: GitHub (org: flagos-ai, 52 repos + 74 commit search results + tag verification), Google News RSS (multiple Chinese and English queries, proxy chain normal), HN Algolia, BAAI Community hub.baai.ac.cn full-text retrieval (see appendix for details)

Index

  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 FlagOS-Compressor tags v0.1.0-rc0.post1: first verification tag for 2.2 new modules lands (09-01)
    • 1.2 community #105: vllm-plugin-FL 0.2 maintenance line added to 2.2 rc0 checklist (09-01)
    • 1.3 build-infra: sglang verification matrix Phase 0 lands, Ascend/MetaX/Moore Threads application image tags registered in bulk (09-01/09-02)
    • 1.4 sglang-plugin-FL: FlagCX KV transfer backend supports PD disaggregation, hcu backend renamed hygon (09-01)
    • 1.5 FlagTree / FlagGems / Torch-FL / FlagCX / FlagFFT and other repo activity (09-01)
  • II. News Coverage and Ecosystem
    • 2.1 Top story: FlagOS three Day0 releases in four days — Qwen3.8-Flash-Next, GLM-5.3-Flash adapted on launch day, Hy4 preview multi-chip release, covering 10 chips (09-01)
    • 2.2 BAAI Community: 2026 embodied intelligence talent migration observations (09-01)
  • III. Member Company Deep Dive
    • 3.1 Enflame: online subscription opens September 2, “domestic GPU four little dragons” assemble in A-shares (09-02)
    • 3.2 Horizon Robotics: H1 revenue 2.055 billion yuan +32.9%, bottom line turns profitable but still loss-making on adjusted basis, licensing model exceeds half (09-01)
    • 3.3 MetaX: joint venture with UBTech to establish Shenzhen Xixuan Chuangzhi, “turning profitable without selling chips” interpretation continues (09-01)
    • 3.4 Hygon: China Everbright Bank procures Hygon CPU platform for 133 million yuan (09-01)
    • 3.5 Other members: Moore Threads appears at HICOOL 2026, Cambricon gets Goldman Sachs target price upgrade, Iluvatar CoreX profitability turnaround interpretation gains traction (09-01)
  • IV. Summary
  • Appendix: Complete Source List

I. Open Source Project Progress (GitHub Activity)

Window Overview: Of the 52 repos in the org, 16 had pushes during the window, and commit search hit 74 in-window commits (build-infra 43, FlagGems 8, sglang-plugin-FL 6, Torch-FL 4, FlagTree 3, etc.). Main threads this window: 1) FlagOS-Compressor tagged v0.1.0-rc0.post1, becoming the second externally validated tag of the 2.2 cycle (following yesterday’s FlagSparse) and the first validation tag to land among the two new modules in the 2.2 manifest; 2) community #105 brings the vllm-plugin-FL 0.2 maintenance line into the 2.2 rc0 release manifest, refining release manifest management granularity down to maintenance branches; 3) the build-infra sglang validation matrix (Phase 0) is formally established, while the Ascend (ascend-cann8.5.0) vLLM application image is registered into unified image management for the first time, and MetaX (metax-maca3.8.1.3/3.7.2.1) and Moore Threads (mthreads-musa5.2.0/4.3.6) image tags are batch-committed; 4) sglang-plugin-FL introduces the FlagCX KV transfer backend to support PD-disaggregated deployment, and the Hygon backend module is renamed from hcu to hygon for consistency.

1.1 FlagOS-Compressor tagged v0.1.0-rc0.post1: first validation tag lands for a new 2.2 module (09-01)

Sources: FlagOS-Compressor tag v0.1.0-rc0.post1 (tagger 2026-09-01 11:30 Beijing time), community #104

  • At 11:30 on 09-01 (Beijing time), FlagOS-Compressor was tagged v0.1.0-rc0.post1 (the tag message notes “FlagOS 2.2 release”); at 11:55 community #104 was merged, formally cutting the rc0 release line from the release/v0.1.0 branch.
  • FlagOS-Compressor is one of the two new modules in the 2.2 RC0 manifest (landed yesterday via #103) — the quantization compression toolchain — and the manifest’s target version is exactly v0.1.0-rc0.post1. This tag aligns fully with the manifest, making it the first component among the new 2.2 modules to complete the “validate → tag” loop.
  • Connecting with yesterday’s FlagSparse v0.3.0-rc0.post1, this is the second externally validated tag of the 2.2 cycle; the repo also has capabilities such as w8a8-moe-export (INT8 MoE weight export) accumulating on the release/v0.1.0 line.

Interpretation: FlagOS-Compressor had just appeared in BAAI’s official Day0 article as the INT8 quantization tooling for the Hy4 preview (see 2.1), and immediately completed rc0.post1 validation and tagging — the quantization compression toolchain has moved from “capability demo” to “version management,” giving the 2.2 INT8 multi-chip release path a formal deliverable.

1.2 community #105: vllm-plugin-FL 0.2 maintenance line added to the 2.2 rc0 manifest (09-01)

Source: community commit d3c1dd5 (#105)

  • At 15:45 on 09-01 (Beijing time), community #105 was merged: adding the release/0.2 maintenance branch line of vllm-plugin-FL to the 2.2 rc0 release manifest. Previously the manifest contained only the main release line entry for vllm-plugin-FL (v0.3.0-rc0.post1); this change brings the 0.2.x maintenance line under version management as well.
  • Repo-side corroboration: both the v0.3.0-rc0 and v0.3.0-rc0.post1 tags for vllm-plugin-FL already exist (rc0.post1 was tagged at 02:03 Beijing time on 09-01, pointing to the #415 CI migration merge commit); #415 on 08-31 had just completed the release/0.2 → main CI code migration, and the two-line parallel structure has taken shape.

Interpretation: The 2.2 manifest is moving from “single-line version locking” to “main version + maintenance line dual-line management,” indicating that vllm-plugin-FL 0.2.x still has a large user base in existing deployments (such as the various registered 0.2.x application images), and the release process must safeguard both new-version validation and old-line maintenance.

1.3 build-infra: sglang validation matrix Phase 0 lands, Ascend/MetaX/Moore Threads application image tags batch-registered (09-01/09-02)

Sources: build-infra #670, #673, #675, #668, #659-#671 series

  • At 19:43 on 09-01, #670 was merged: sglang validation matrix Phase 0 + backend validation scripts — the status-matrix expanded from vLLM/megatron to the sglang line, and at 20:24 Phase 0 was marked closed (#670 merged). This is the third validation matrix formally established, following the megatron dual-backend matrix and the vLLM multi-vendor matrix.
  • Ascend vLLM application image registered for the first time: at 13:06 the 2.1.2-0.2.1 application image tag for ascend-cann8.5.0 was registered — Huawei Ascend joins FlagOS unified vLLM application image management, corroborating the Huawei Ascend deployment on the SGLang-plugin-FL path in the 2.1 Day0 article.
  • MetaX dual-version registration: at 20:45/22:15 the vLLM 0.2.1 image tags for the metax-maca3.8.1.3 and metax-maca3.7.2.1 backends were registered; at 23:34 #675 recorded metax PR #377 into the vLLM 0.24.0 status matrix.
  • Moore Threads registration: at 15:02/22:25 the vLLM 0.3.0rc0 application image tags for mthreads-musa5.2.0 and musa4.3.6 were registered (continuing the rc line built on 08-17); the nvidia-cuda13.3 image was updated in sync.
  • At 22:56 #673: megatron wheel build adds NO_VCS_VERSION=1 (eliminating the version derivation’s dependency on VCS information); at 17:35 #668 fixed GH_TOKEN passing in the status-matrix CI rendering step; from 14:30 onward, multiple commits hyperlinked status-matrix PR numbers (clickable straight to the PR).

Interpretation: The theme for build-infra this window is “validation matrix engineering” — sglang line Phase 0 establishment + real-time PR status column rendering + pipelined multi-vendor image tag registration. Together with yesterday’s megatron dual-compiler-path validation, both the training and inference validation lines are converging toward “multi-vendor, traceable, semi-automated”; Ascend entering unified image management for the first time signals expanding member coverage.

1.4 sglang-plugin-FL: FlagCX KV transfer backend supports PD disaggregation, hcu backend renamed hygon (09-01)

Sources: sglang-plugin-FL #59, #78, #63, #40, #80

  • At 20:24 on 09-01, #59 was merged: adding the FlagCX KV transfer backend to support PD (Prefill/Decode) disaggregated deployment — bringing the FlagCX cross-chip communication library into the KV transfer path at the sglang plugin layer, so cross-node KV migration in PD-disaggregated scenarios no longer depends on a single vendor’s communication implementation. This echoes the flagcx v0.14.0-rc0.post1 target version in the 2.2 manifest.
  • At 19:29, #78: the hcu backend module path was renamed to hygon — standardizing Hygon (Hygon DCU) backend naming, aligned with the “Hygon DCU” wording in the Day0 article.
  • At 19:53, #63: PPU default attention_backend registered into _ATTN_BACKEND_MAP; at 16:29 #40 added an end-to-end throughput benchmark script for the sglang server side; at 16:32 #80 fixed multi-node testing (full validation of prime sequences); at 16:33 #69 aligned SGLANG_FL_STRICT semantics.

Interpretation: sglang-plugin-FL is advancing simultaneously on three lines — “PD disaggregation (FlagCX KV transfer) + backend naming standardization (hygon) + benchmarking tooling” — and the plugin layer is moving from “it runs” to “testable, disaggregation-deployable”; FlagCX’s role on the inference side is expanding from a training communication library to PD-decoupled transfer, widening the value radius of the cross-chip communication stack.

1.5 Other repo activity: FlagTree / FlagGems / Torch-FL / FlagCX / FlagFFT (09-01)

Sources: FlagTree #1055, #1035, FlagGems commits, Torch-FL #237/#239/#242/#243, FlagCX #559, FlagFFT commits

  • FlagTree: at 14:33 #1055 FlagTune supports more accelerator backends; at 21:21 #1035 FlagPrism’s Iluvatar integration was migrated and formatted onto the main branch; at 12:21 #1077 upgraded the NVIDIA qwen benchmark to vLLM 0.24.
  • FlagGems (8 commits): at 15:42 #5816 kunlunxin backend slice_scatter out-of-bounds (OOB) fix — the Kunlunxin backend remains actively maintained in the operator library (corroborating Kunlunxin in the Day0 article’s 10-chip list); at 18:45 #5876 NVIDIA Hopper MM split-K accumulation precision fix; 4 performance optimizations on the MTHREADS side (upsample_linear1d_backward 2D grid, conv_transpose1d, channel_shuffle block-copy, conj); at 18:09 #5885 version derivation tolerates post-release tags (paired with the rc0.postN tagging process); at 11:28 rule-check CI phase1 continued to advance.
  • Torch-FL: at 11:57 #237 renamed the HF survey skill to transformers-test and added an all-architecture sweep — landing the methodology accompanying yesterday’s #236 official test runner; at 13:57 #239 fixed the MUSA backend binary operators’ fault on CPU wrapped scalars; at 16:43 #242 rewrote degenerate size-1 dimension strides before handing tensors to mudnn; at 23:24 #243 added a reusable test-dependency skill.
  • FlagCX: at 14:40 #559 adapted PAL to the new FlagCX interface (MetaX backend).
  • FlagFFT: 11:01-11:08 the MUSA S5000 (Moore Threads) FP64 packed path passed validation and was enabled, FP64 Bluestein fusion, and backend-specific plan gating restricted by device type.
  • vllm-plugin-FL: at 10:35 #426 supports custom throughput test cases; FlagGems-vllm: triple update to pytest markers/tests/benchmark/CI.

Interpretation: The Kunlunxin backend (FlagGems #5816), the Iluvatar FlagPrism integration (FlagTree #1035), and the Day0 article’s 10-chip landscape corroborate one another, showing member chip backends deepening their footprint across both the operator library and compiler layers; Torch-FL’s transformers-test all-architecture sweep solidifies yesterday’s “official test runner” methodology into a reusable skill asset.

II. News and Ecosystem

2.1 Headline: FlagOS Delivers Three Day0 Releases in Four Days — Qwen3.8-Flash-Next, GLM-5.3-Flash Adapted on Launch Day, Hy4 preview Multi-Chip Release, Covering 10 Chips (09-01)

Sources: BAAI Official Release (hub.baai.ac.cn/view/57616) (2026-09-01 20:10), vLLM upstream PR 53909, vLLM upstream PR 53877

  • On the evening of 8/26, Alibaba open-sourced Qwen3.8-Flash-Next and Zhipu open-sourced GLM-5.3-Flash; on the afternoon of 8/27, Tencent open-sourced Hy4 preview. The Zhongzhi FlagOS community completed three consecutive rounds of multi-chip adaptation: both Qwen3.8-Flash-Next and GLM-5.3-Flash achieved Day0 on launch day, with Hy4 preview subsequently completing its multi-chip release; covering 10 chips including T-Head, NVIDIA, Moore Threads, Huawei Ascend, MetaX, Kunlunxin, Hygon, Iluvatar CoreX, Tsingmicro, and Xiwang, with BF16/FP8/INT8 versions provided according to hardware capabilities, progressively released to ModelScope and Hugging Face.
  • Qwen3.8-Flash-Next (early preview of Qwen4 architecture): Introduces three new structures — QSA sparse attention, Gated Residual, and N-gram Embedding. FlagOS added 8 new Triton-based operators (3 Hyper-Connection, 3 QSA, 2 PLE state) and compiled them to multiple chips via FlagTree; operator fusion saves approximately 8.66ms per step (HC fusion contributes 6.3ms, accounting for over 70%); the key was incorporating metadata construction into CUDA Graph, going from partial graph capture to full graph capture yielding +7.46% end-to-end gains (only +0.93% at the individual operator level). The 8 fused operators have been submitted to vLLM upstream (PR 53909), along with an upstream fix for an FP32 precision defect in the GDN Packed Decode path (PR 53877).
  • GLM-5.3-Flash (first native multimodal model in the GLM-5 series, AA composite intelligence index 57): Parallel partitioning differences of linear+sparse hybrid attention across different chips are handled by the inference plugin layer; the vLLM-plugin-FL and SGLang-plugin-FL paths share the same set of FlagGems operators and BF16 weights, requiring no model file changes when switching chips.
  • Hy4 preview (770B parameters, 2.6× Hy3, 49B activated, 1M context): Achieves an INT8 compression path from native FP8/BF16 weights via the FlagOS-Compressor quantization toolchain, and completes multi-chip adaptation of quantized inference operators; after quantization + cross-chip migration, errors compared against the original NVIDIA CUDA FP8 version fall within the alignment interval.
  • Release system data update: As of August 2026, FlagRelease has published cross-chip versions covering 10+ chip vendors, 12+ hardware platforms, and 80+ open-source model instances.

Analysis: This follows the 8/13 Qwen3.8-2.4T nine-chip Day0 and the April DeepSeek-V4 eight-chip Day0, marking the first time FlagOS has completed frontier model multi-chip adaptation at a “three consecutive Day0” density — officially attributed to the four-layer compounding of the public operator library, unified compiler (FlagTree), compression tool (FlagOS-Compressor), and standardized release pipeline (FlagRelease), signaling that cross-chip adaptation is moving from project-based to infrastructure capability; the bidirectional vLLM upstream contributions (operators + precision fixes) demonstrate that FlagOS adaptation results are being fed back into mainstream international inference frameworks.

2.2 BAAI Community: 2026 Embodied Intelligence Talent Migration Observation (09-01)

Source: BAAI Community “2026 Embodied Intelligence Talent Migration Observation” (BAAI Community, 09-01)

  • The BAAI Community published the “2026 Embodied Intelligence Talent Migration Observation,” tracking talent flows in the embodied intelligence sector (migration between universities/big tech/startups, changes in talent composition at embodied intelligence companies, etc.), continuing BAAI’s research output in the embodied intelligence direction (part of the same content line as yesterday’s “BAAI Scholars Embodied Intelligence Workshop” and the BAAI Conference agenda).

Analysis: BAAI maintains a multi-dimensional layout in embodied intelligence — “academic workshops + conference agenda + talent research” — consistently echoing the FlagOS-Robo embodied intelligence toolchain direction; talent migration observation content can serve as a side signal of FlagOS-Robo ecosystem activity.

III. Member Unit Deep Dive

3.1 Enflame: Online subscription opens September 2, the “Four Little Dragons of Domestic GPUs” assemble on the A-share market (09-02)

Sources: China Securities Network (subscription announcement), Phoenix Finance (one winning lot may earn over 280,000 yuan), Shanghai Securities News (chairman speaks out), China News Jingwei (investor exchange meeting)

  • On September 2 (today), Enflame Technology (688801.SH) opens online and offline subscription, at an issue price of 142.18 yuan/share (reported yesterday), with one winning lot of 500 shares corresponding to a payment of 71,100 yuan; media, based on the first-day performance of recent hundred-yuan IPOs, estimate that “one winning lot may earn over 280,000 yuan,” calling it a “big meat lot”; the “Four Little Dragons of Domestic GPUs” (Cambricon, Hygon, MetaX, Enflame) have now all assembled on the A-share market.
  • Latest statement from Chairman Zhao Lidong (9/1 roadshow + investor exchange meeting): computing power demand is growing explosively, and the company’s revenue for the first three quarters is expected to grow 325.78% to 455.36% year-on-year, with profitability expected by 2027 at the latest; the IPO online roadshow was extremely popular (160 questions in 3 hours), with 337 institutions and 11,805 allocation targets participating in the offline inquiry, and an overall subscription multiple of 2,648.93x (reported yesterday).

Interpretation: Enflame’s IPO has entered the stage of new-share subscription and cashing out. After the “Four Little Dragons” meet in the capital market, the domestic AI chip sector moves from “comparing revenue” to “comparing delivery and profit realization” (Sohu’s headline is precisely “Domestic AI chips begin competing on delivery capability”); its collaboration with the FlagOS GCU stack (FlagCX/FlagTree gcu backend) remains a point to watch going forward.

3.2 Horizon Robotics: H1 revenue 2.055 billion yuan, +32.9%, turned profitable on the statements but still loss-making on an adjusted basis, licensing model accounts for more than half (09-01)

Sources: Chedongxi (interim report breakdown), Sina Finance (Yu Kai’s interpretation), Sing Tao Global Network (Journey 6M enters mass production and hits the road), Caixin Global (adjusted loss widens)

  • On the evening of 8/31, Horizon Robotics (09660.HK) released its 2026 interim report: revenue 2.055 billion yuan (+32.9%), gross margin 66.0%; net profit on the statements was 3.784 billion yuan (versus -5.233 billion yuan in the same period of 2025), achieving a turnaround, but mainly from gains related to the deconsolidation of D-Robotics and changes in the fair value of financial liabilities and other non-recurring items; adjusted net loss was 1.671 billion yuan, widening 25.4% year-on-year (Caixin says the founder aims to achieve break-even by 2028).
  • Signal of a qualitative change in the business model: revenue from licensing and services rose to 55% (+7.2pct), while product solutions fell to 45%, “the licensing model has entered the realization stage”; cumulative design wins for nearly 500 vehicle models (nearly 130 for mid-to-high-end intelligent driving); Journey 6M officially announced mass production and vehicle deployment, with urban NOA extending down to the 100,000-yuan-class market.
  • Product roadmap: development of the next-generation strategic SoC Journey 7 is progressing smoothly, with tape-out expected in early Q2 2027 and market launch in 2027; the cockpit-driving integrated solution based on the Starry Sky chip (cockpit-driving fusion) will enter mass production in Q4 2026; HSD has already won design wins with top-five Chinese self-owned brands by sales and will begin delivery in the second half of the year; L3 capabilities of the Volkswagen joint venture CARIZON will begin delivery in H2 2027.
  • Yu Kai stated: confident that full-year revenue will exceed 5 billion yuan; declared that “starting in 2027, Horizon Robotics’ high-level intelligent driving chips and IP-licensed automakers’ self-developed chips will achieve the number one market share in China’s high-level intelligent driving chip market”; overseas expansion has already covered 24 brands and nearly 60 vehicle models.

Interpretation: As a FlagOS member unit (D-Robotics BPU), Horizon Robotics’ “profitable on the statements but still loss-making on an adjusted basis” and the discussion of the “quality of turnaround” at MetaX/Iluvatar CoreX formed the main line of financial media on 8/31-9/1 (“Among domestic GPU/chip makers, whose main business is making money”); the licensing model accounting for more than half means its collaboration with the FlagOS open-source software stack (BPU backend adaptation) has stronger business model support.

3.3 MetaX: Joint venture with UBTech to establish Shenzhen Xixuan Chuangzhi, interpretation of “turnaround not relying on selling chips” continues (09-01)

Sources: Sina Finance (establishment of Xixuan Chuangzhi), Sina Finance (turnaround interpretation), The Paper (behind the high revenue growth of domestic GPUs)

  • On 09-01, business registration information showed: MetaX and a company held by UBTech formed a joint venture in Shenzhen called “Xixuan Chuangzhi Technology Company”—an equity-binding move between a GPU maker and a leading embodied intelligence company (UBTech), continuing MetaX’s layout in the embodied intelligence direction.
  • Media continued to break down MetaX’s interim report: “Behind the turnaround to profitability, the comeback is not from selling chips” (revenue 1.324 billion yuan, +44.67%, but non-operating gains accounted for a high proportion of the 612 million yuan net profit; as reported yesterday, it still lost 48.86 million yuan on a non-GAAP basis); The Paper’s “Behind the high revenue growth of domestic GPUs in the first half: who is making money in their main business, and who is turning around through wealth management” listed it alongside Iluvatar CoreX as a sample.

Interpretation: Within FlagOS, MetaX remains active on the MUSA backend (FlagGems MTHREADS optimization, Torch-FL #239/#242, FlagFFT S5000 FP64), with ecosystem-side progress (Xixuan Chuangzhi + Day0 coverage) proceeding in parallel with financial-side “quality discussion”; the joint venture with UBTech points toward robotics/embodied intelligence computing scenarios.

3.4 Hygon: China Everbright Bank procures Hygon CPU platform for 133 million yuan (09-01)

Sources: Drivers Home (China Everbright Bank procurement)

  • 09-01 report: China Everbright Bank procured a Hygon CPU platform for 133 million yuan (bank AI computing power/xinchuang infrastructure), with media calling it a “vote with real money,” as the Hygon CPU platform captured a core position in bank AI computing power.
  • Hygon remained highly active in this window: as reported yesterday, it won the China Patent Gold Award and jointly adapted with Tencent Cloud Cube Sandbox; on the FlagOS side, sglang-plugin-FL #78 standardized backend naming (hcu → hygon), and in the Day0 article Hygon DCU covered three models: Qwen3.8-Flash-Next, GLM-5.3-Flash, and Hy4 preview.

Interpretation: Hygon’s three lines—”patent gold award + major bank order + triple Day0 coverage”—resonate together, making it one of the most balanced vendors among recent FlagOS members in both ecosystem and commercialization; its official DCU PyTorch route (Torch-FL) and collaboration with FlagOS continue to benefit from this level of activity.

3.5 Other members: Moore Threads appears at HICOOL 2026, Cambricon gets target price raised by Goldman Sachs, Iluvatar CoreX turnaround interpretation fermenting (09-01)

Sources: Sanyi Life (Moore Threads at HICOOL), Sina Finance (Goldman Sachs raises Cambricon target price), East Money (Iluvatar CoreX)

  • Moore Threads: appeared at the HICOOL 2026 Global Entrepreneur Summit (Sanyi Life, 09-01); on the FlagOS side, the MTHREADS backend and MUSA line saw intensive optimization commits in this window (4 FlagGems items, FlagFFT S5000 FP64 verification, Torch-FL MUSA fixes), while also registering two backend vLLM 0.3.0rc0 application images, mthreads-musa5.2.0/4.3.6.
  • Cambricon: Goldman Sachs raised its target price (09-01 report), and “how to grasp the Cambricon moment for domestic chips” became a discussion topic; on the FlagOS side, Cambricon is one of the ten Day0 chips (Qwen3.8-Flash-Next/GLM-5.3-Flash).
  • Iluvatar CoreX: the interpretation that it “went from a loss of 600 million to a profit of 100 million, turning around through stock trading gains” continued to ferment (East Money); on the FlagOS side, the Iluvatar backend saw substantive progress in this window, including FlagTree FlagPrism integration migration (#1035) and megatron corex4.5.0 verification (yesterday).

IV. Summary

In this window (09-01 10:18 ~ 09-02 10:18), the main lines of the FlagOS ecosystem:

  1. BAAI’s official release of “FlagOS three Day0s in four days” was the biggest event of this window: same-day adaptation for the releases of Qwen3.8-Flash-Next and GLM-5.3-Flash, and multi-chip release for Hy4 preview, covering 10 chips; the 8 fused operators and 1 precision fix for Qwen3.8-Flash-Next have already been contributed back upstream to vLLM; the INT8 implementation of Hy4 preview relies on the FlagOS-Compressor quantization toolchain—”continuous Day0” is officially defined as the compounding effect of a public software stack.
  2. 2.2 RC0 verification advances: FlagOS-Compressor tagged v0.1.0-rc0.post1 (the first verification tag for a newly added 2.2 module, fully aligned with the manifest), community #105 brought the vllm-plugin-FL 0.2 maintenance line into the 2.2 manifest, and release management granularity was refined down to maintenance branches.
  3. Engineering of the build-infra verification matrix: the sglang verification matrix Phase 0 was initiated and closed (the third matrix after vLLM/megatron), the Ascend (ascend-cann8.5.0) vLLM application image was registered into unified management for the first time, MetaX/Moore Threads image tags were batch-entered into the repository, and the PR status column was launched with real-time rendering.
  4. sglang-plugin-FL introduced the FlagCX KV transfer backend to support PD-disaggregated deployment, and Hygon backend naming was standardized (hcu → hygon); FlagCX expanded from a training communication library to inference-disaggregated deployment scenarios.
  5. Dense member unit activity: Enflame opened online subscription today (9/2), with the “Four Little Dragons” assembling on the A-share market (the chairman says profitability by 2027 at the latest); Horizon Robotics’ H1 revenue +32.9%, profitable on the statements but still loss-making on an adjusted basis, licensing model accounting for more than half, and Journey 7 scheduled for tape-out in 2027Q2; MetaX formed the Xixuan Chuangzhi joint venture with UBTech; Hygon won a 133 million yuan procurement from China Everbright Bank.
  6. Component-level queries on the news side still mainly yield scattered hits (the only direct hit in this window was BAAI’s official Day0 article); GitHub commits and member unit financial developments remain reliable sources of activity, and the component version cadence (rc0.postN verification tags) has entered a period of intensive implementation.

Appendix: Complete Source List

Source Verification Result
GitHub org repos API (52 repos) 16 repos pushed within window, all 74 commits verified
GitHub commit search 74 commits within window (all 3 pages retrieved), no omissions
GitHub tags API (FlagOS-Compressor / vllm-plugin-FL) Tag objects verified: flagos-compressor v0.1.0-rc0.post1 (tagged 09-01 11:30 CST), vllm-plugin-FL v0.3.0-rc0.post1 (09-01 02:03 CST, pointing to #415); no new releases in this window per releases API
community #104/#105 Verified item by item: flagos-compressor rc0 branch cut, vllm-plugin-FL release/0.2 line merged into 2.2 rc0
build-infra #659-#675 series sglang Phase 0 (#670), Ascend cann8.5.0 first registration, MetaX/Moore Threads image tags, #673 wheel fix, #675 metax PR record verified item by item
sglang-plugin-FL #59/#78/#63/#40/#80/#69 Verified item by item: FlagCX KV transfer PD disaggregation, hcu→hygon, PPU attention backend, throughput benchmark scripts
FlagTree/FlagGems/Torch-FL/FlagCX/FlagFFT/vllm-plugin-FL/FlagGems-vllm Commit titles and PR statuses verified; FlagGems kunlunxin #5816, Hopper #5876, etc. confirmed item by item
BAAI community hub.baai.ac.cn/view/57616 Full original text retrieved and verified: three Day0 events in four days, 10-chip list, 8 operators, CUDA Graph +7.46%, vLLM upstream PRs 53909/53877, FlagRelease 80+ model instances
Google News RSS Chinese group 1 hit for FlagOS component-level queries (the BAAI official Day0 article); multiple items verified for Enflame (today’s subscription/signing gains/Zhao Lidong’s remarks), Horizon Robotics (interim report/Journey 6M/Yu Kai), MetaX (Xixuan Chuangzhi/loss-reversal interpretation), Hygon (Everbright Bank 133 million), Moore Threads (HICOOL), Cambricon (Goldman Sachs target price)
Google News RSS English group No valid hits (Korean media Chosunbiz report on domestic GPU cash flow as industry background)
HN Algolia Zero relevant hits for FlagOS/FlagGems/FlagScale/FlagTree/FlagAttention; 8 hits for BAAI queries, none directly related to FlagOS
CheDongXi (Horizon Robotics interim report) Body text verified: 2.055 billion +32.9%, net profit 3.784 billion (including gains from terminating D-Robotics merger), adjusted loss 1.671 billion, licensing share 55%, Journey 7 tape-out in 2027Q2, Starry Sky mass production in Q4
CS.com.cn/Phoenix/Sina/Shanghai Securities News/China News Service (Enflame) Subscription on 9/2, 688801.SH, one winning lot may earn over 280,000, first three quarters revenue forecast +325.78%-455.36%, profitability by 2027 at the latest, verified item by item
Sina Finance/The Paper (MetaX, Tianshu Zhixin) Xixuan Chuangzhi established, “loss reversal not reliant on selling chips,” discussion of domestic GPU profitability quality verified