Research Window: 2026-09-04 10:18 ~ 2026-09-05 10:18 Beijing Time Sources: GitHub (org: flagos-ai 52 repos + full commit search 51 entries + per-repo default branch verification + tags/releases date verification), Google News RSS (21 query sets in Chinese and English, proxy chain operational), HN Algolia, financial media original-site retrieval and verification (Sina Tech/Operator Finance/Times Finance/Lianhe Zaobao/Kuaikeji, etc., see appendix)

Index

  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 FlagTree integrates Tsingmicro backend for the first time: 300 files land TLE-DSA dataflow abstraction (09-04)
    • 1.2 build-infra: vLLM 0.24.0 validation matrix expands — Cambricon passes, Iluvatar T path deemed undeliverable (09-04)
    • 1.3 build-infra: sglang 0.5.18 application environment lands on Cambricon and Enflame, dense merges of image records and build fixes (09-04/09-05)
    • 1.4 FlagGems-vllm: Ascend operators batch-completed and cann850 removed, Hygon/Moore Threads normalization operator optimization (09-04/09-05)
    • 1.5 FlagGems: KMCompiler new operators and FlagTune cost model, W4A8 quantized linear operators on ARM CPU side (09-04/09-05)
    • 1.6 Other components: FlagDNN adds Moore Threads backend, vllm-plugin-FL scheduling and quantization fixes, FlagScale dual commits (09-04)
  • II. News Coverage and Ecosystem
    • 2.1 Component-level news quiet for a fourth consecutive window, release lull continues (09-04~09-05)
    • 2.2 BAAI institutional activity: no FlagOS-related official releases; “智源新境” is a Shanghai marketing technology company, name collision with BAAI removed (09-04)
  • III. Member Company Deep Dive
    • 3.1 Enflame: IPO pricing set at 142.18 yuan/share, market cap ~61.2 billion yuan, H1 loss widened to 632 million yuan, roadshow provides 2026/2027 profitability timeline (09-04)
    • 3.2 Moore Threads: “Lushan” GPU to launch within the year, first domestic hardware ray tracing; H1 revenue +147% with losses significantly narrowed; PD heterogeneous collaboration with Qujing Technology (09-04/09-05)
    • 3.3 MetaX: book net profit of 612 million yuan actually relies on 887 million yuan in unrealized gains, non-GAAP still loses 48.86 million yuan, mainstream media collectively dissect “profit turnaround quality” (09-04)
    • 3.4 Iluvatar: book profit turnaround of 106 million yuan entirely propped up by Shenghe Jingwei unrealized gains, gross margin plunges to 17.2%, core business losses widen (09-04)
    • 3.5 Cambricon: 11 teams aggressively recruiting software talent, doubling down on AI software ecosystem (published 09-02/covered 09-04)
  • IV. Summary
  • Appendix: Complete Source List

1. Open-Source Project Progress (GitHub Activity)

Window Overview: Of the 52 repos in the org, 11 had pushes during the window, and commit search returned 51 commits within the window (build-infra 19, FlagGems 11, FlagGems-vllm 9, FlagTree 5, vllm-plugin-FL 3, FlagDNN 2, FlagScale 2); pushed_at changes for docs/release-info/flagtree-cpu/sglang-plugin-FL were verified via default-branch commits to be non-default-branch (PR branch) activity, with no substantive merges within the window. Verification of tags/releases dates (including annotated tag objects) shows no new tags and no new releases within the window — the two 2.2 RC0 validation-period tags, vllm-plugin-FL v0.3.0-rc0.post1 and FlagGems-vllm v0.2.0-rc0.post1, were both created on 08-31 (UTC), continuing the stable release period. Main threads of this window: 1. FlagTree integrates the Tsingmicro backend for the first time at a 300-file scale, adding the TLE-DSA (dataflow architecture) programming abstraction and DsaToCore conversion, formally bringing Tsingmicro into the unified compiler’s backend matrix; 2. build-infra’s vLLM 0.24.0 validation matrix expands again — Cambricon and Iluvatar (iluvatar-corex4.4.0) application images pass validation, but the Iluvatar T path (vendor Triton 3.1.0) is formally judged “undeliverable” due to an engine initialization defect, marking the first divergence in the multi-vendor delivery matrix where “the F path is deliverable while the T path is not”; 3. sglang 0.5.18 application environments (deps_app/app env) land in batches for Cambricon and Enflame, with validation records and build fixes densely merged (12 of the 19 build-infra commits in the window belong to the sglang line), pushing the sglang image matrix into a rolling “register-validate-fix” state.

1.1 FlagTree Integrates the Tsingmicro Backend for the First Time: 300 Files Land the TLE-DSA Dataflow Abstraction (09-04)

Source: FlagTree #1071 (4f61a9d8ca)

  • 09-04 21:02 (Beijing time) #1071 merged: adds Tsingmicro backend integration (Triton 3.6) to FlagTree — this commit touches 300 files, the largest backend merge in the window and the first substantive Tsingmicro support in FlagTree’s history (previously only a documentation-level mention on 07-16, verified via commit search).
  • Code structure: adds the python/triton/experimental/tle/language/dsa/ submodule (including dsa/core.py, dsa/semantic.py, dsa/types.py, and the dsa/tsingmicro/ vendor implementation), and adds the tle-dsa dialect (DsaDialect/DsaOps.td) and the DsaToCore conversion under third_party/tle; accompanied by the setup_tools/utils/tsingmicro.py build script, the UnifiedHardware abstraction, and a full set of TLE DSA integration tests (test_tle_dsa_arith/bridge/pipeline_e2e/rand/slice, bench_tle_dsa.py, test_tle_cumsum_txda.py).
  • Semantic highlights: DSA (Dataflow Streaming Architecture) is the programming abstraction for Tsingmicro’s reconfigurable dataflow chips — by adding DSA primitives at the TLE language layer and performing the DsaToCore lowering inside the compiler, Tsingmicro’s “dataflow execution” model is brought into the Triton/TLE ecosystem.

Interpretation: Tsingmicro is the most unusual player among FlagOS member organizations (reconfigurable/dataflow architecture, distinct from many-core GPGPUs), and had previously been progressively integrated at the FlagScale (txda platform, 07-28), FlagGems (regular TSINGMICRO backend updates), and build-infra (08-31 registration of tsingmicro-tsm260610’s vLLM 0.24.0 deps_app) levels. This 300-file-scale FlagTree merge completes the “unified compiler” link: Tsingmicro becomes yet another member to enter the FlagTree backend after Ascend, Cambricon, Iluvatar, Enflame, Moore Threads, Hygon, and MetaX, and it does so via an independent DSA dialect rather than “simulating a many-core”, indicating that the compiler layer preserves a native expression path for reconfigurable architectures. The Triton 3.6 alignment is also consistent with the earlier mthreads3.6 toolchain cadence.

1.2 build-infra: vLLM 0.24.0 Validation Matrix Expands — Cambricon Passes, Iluvatar T Path Judged Undeliverable (09-04)

Source: build-infra #718 (37cc0a5e0c), #722 (d116a8cbd2), #724 (34c0d16e4d)

  • 09-04 11:49 #718: marks Cambricon and Iluvatar (iluvatar-corex4.4.0) vLLM 0.24.0 application images as validated — following Enflame tops1.10.6’s completion of F+T dual-path delivery on 09-03, the 0.24.0 delivery matrix expands to two new lines, Cambricon and Iluvatar.
  • 09-04 15:32 #722 (a record-type commit, the most informative): Iluvatar iluvatar-corex4.4.0’s 0.24.0 T path judged “undeliverable” — reproduced and verified on an ix15 machine (corex 4.4.0 SDK, image g84a4ca2): the T path (vendor Triton 3.1.0) stably reproduces an engine initialization failure during profile_run, with the root cause being a uint32/int32 mixed division in vLLM 0.24.0’s apply_top_k_top_p_triton ternary search kernel, which the corex 3.1.0 branch rejects (same origin as the toolchain-generation-gap issues recorded in the 0.20.2 era); the plugin’s existing iluvatar sampler patch only covers compiled_random_sample and not apply_top_k_top_p, and since EngineCore is a spawn subprocess, parent-process runtime patches cannot propagate — the fix direction is explicitly stated as “must force a pytorch sampler fallback within the plugin module scope”.
  • 09-04 17:42 #724: registers the 09-04 re-verification record for Cambricon 4.7.2 (vLLM 0.20.2) already-released application images — the post-release re-verification (reverify) mechanism for older-version images continues.

Interpretation: This is the first clear divergence record in the multi-vendor delivery matrix where “the F path (FlagGems/KMCompiler generation path) is deliverable while the T path (vendor Triton path) is not” — for the same chip and the same image, the delivery status of the two compilation paths is explicitly forked, and the reason for undeliverability is precisely localized to the vendor toolchain’s (corex 3.1.0) rejection of a specific operation and a coverage gap in the plugin’s sampler patch. This kind of “negative record” (documenting why delivery is impossible plus giving a fix path) is itself a manifestation of engineering credibility: Iluvatar’s 0.24.0 delivery currently stands on the F path (already validated in #718), while the T path awaits a plugin-side sampler fallback patch before re-validation.

1.3 build-infra: sglang 0.5.18 Application Environments Land for Cambricon and Enflame, Image Records and Build Fixes Densely Merged (09-04/09-05)

Source: build-infra #719 (4b8806bc07), #720 (29be2deb3e), #721 (e4950a66ce), #723 (4e98245456), #725 (84d93933f0), #726/#727/#728, #729 (81afd7cea6), #730 (110401a708), #731 (69bff171a3), #732 (1bfa07bade), #733 (07da65b40d), #734 (393649df1e), #735 (46063c3d00), #736 (c64e924901)

  • Application environment (app env) expansion: 09-04 22:10 #729 adds the sglang 0.5.18 deps_app key and application environment for Enflame; 22:20 #730 lands the 0.5.18 application environment for Cambricon as well — the multi-vendor image build matrix for sglang 0.5.18 is now fully rolled out (the MetaX and Ascend cann9.0.0 sglang image tags registered earlier in the window belong to the same 0.5.18 version line). 09-05 08:06 #734 registers the Enflame enflame-tops1.10.6 sglang application image tag (2.1.2-0.1.dev1_g65dff2733).
  • Status and validation records: 09-04 20:51/21:16 #726/#727 register the Cambricon and Enflame plugin PRs in the sglang status matrix; 21:37 #728 appends Cambricon/Enflame validation records and notes Nvidia as pending; on the 11:49 side, the vllm line #718 advances in sync with the sglang line.
  • Dense build and validation fixes (12 commits): #719 publishes to all app-served vendor indexes after shim upload; #720/#732 deduplicate –network host in the wheel build container; #721 routes cargo through the rsproxy.cn domestic mirror; #723/#731 add torchaudio/torchvision to the runtime dependencies for Cambricon and Moore Threads (mthreads); #725 pins compressed-tensors in the Cambricon deps_app; #733 changes the serve readiness gate to wait for the post-warmup log line; #735 closes on verify failure on the Enflame side and bakes in a warmup timeout; #736 fixes bare-quote escaping in the Step 7 exec string.

Interpretation: build-infra’s sglang line has moved from “registering image tags” into a construction phase of “completing application environments for vendors + runtime fixes”: the Cambricon and Enflame sglang 0.5.18 images are about to enter a validation-delivery process at the same specification as vLLM 0.24.0; the fixes concentrate on wheel build networking (rsproxy mirroring), container network deduplication, dependency completion, and readiness gating — all engineering-foundation actions that make “building vendor images repeatably in CI” possible. Worth noting is the mechanism in #719: once a shim is uploaded it is automatically published to all app-served vendor indexes, indicating a rising degree of automation in image distribution.

1.4 FlagGems-vllm: Ascend Operators Filled In at Scale and cann850 Removed, Hygon/Moore Threads Normalization Operator Optimizations (09-04/09-05)

Source: FlagGems-vllm #731 (e455c57d06), #728 (9953ab4f6b), #729 (968d832a1b), #730 (d167d8f008), #740 (255ab5601b), #738 (5865ed53b1), #735 (5643b85a7c), #697 (98aec3d6bb)

  • Ascend operator fill-in at scale: 09-04 16:05~17:25 saw four consecutive Ascend backend operator merges — add_rms_norm (#731), swiglu (#728), causal_conv1d_update (#729), causal_conv1d_fn (#730, including Ascend unit tests); 09-05 01:41 #740 removes the cann850 backend on the tools side — synchronized with yesterday’s FlagGems #5970 removal of cann850, the Ascend toolchain completes its move away from CANN 8.5 in both the operator library and the bridge layer.
  • Normalization operator optimizations: 09-04 15:07 #738 optimizes the Hygon backend’s fused QKV RMSNorm (a vLLM inference hot path); 14:58 #735 optimizes Moore Threads’ (mthreads) fused add rms norm. In the same batch, #697 adds fp8/fp4 MQA logits operators to KMCompiler TLE (presumably quantized MQA attention logits generation).
  • Engineering fixes: 09-05 01:14 #739 fixes a circular import error between triton and torch.

Interpretation: The 9 commits in FlagGems-vllm during the window show two clear lines: on the Ascend side, “fill in operators + switch toolchains” (four new operators completing the swiglu/RMSNorm/causal convolution family — causal convolution is a core operator for Mamba-class SSM models, indicating the Ascend backend is covering non-Transformer architectures; the cann850 removal corroborates the CANN 9.0.0 registration on the build-infra side) and domestic-vendor inference hot-path optimization (the RMSNorm variant optimizations for Hygon/Moore Threads show that vLLM inference performance tuning for each backend has reached the operator level). Although Ascend is not a signed FlagOS member organization, the depth of its toolchain and operator adaptation is already on par with member organizations.

1.5 FlagGems: New KMCompiler Operators and FlagTune Cost Model, W4A8 Quantized Linear Operator on the ARM CPU Side (09-04/09-05)

Source: FlagGems #6002 (7e6c5524f2), #5942 (156e7e475f), #5958 (5bedd841e7), #5762 (8416e6dab4), #5904 (cd55e57b42), #5980 (9553ae9c74), #6004 (855774a3ed), #5999 (a3d2fc8c58)

  • KMCompiler (NVIDIA generation path) operators and tests: 09-04 19:26 #6002 adds a gru operator with a Triton kernel; 09-05 03:22 #5942 adds a matrix rank operator; 15:30 #5958 refines the distributed operator implementation and completes full test coverage — KMCompiler’s operator surface extends from elementwise/matrix multiplication toward RNN (GRU) and linear algebra (rank).
  • FlagTune cost model: 09-05 03:47 #5762 adds cost model support for the multi-platform Mul (multiplication) operator to FlagTune — the tuner’s cost modeling extends to multi-chip platforms.
  • New capabilities on the ARM CPU side: 09-04 14:54 #5904 adds a W4A8-G128 quantized linear API to the CPU backend (Triton CPU implementation, ARM platform); 09-04 14:56 #5999 fixes the uv unzip path issue on aarch64 in setup.sh — combined with the flagtree-cpu repo (which had an ARM64 CPU backend merge on 07-18) and yesterday’s [ARM] series of commits, FlagGems’ quantization support for the ARM/CPU inference path continues to be filled in.
  • Fix-type commits: 09-04 11:01 #5980 fixes NaN/OOM and the cumsum device issue in the Ascend backend’s multinomial; 09-05 00:15 #6004 fixes slice’s missing default arguments and view semantics breakage; 16:46 #5930 fixes special_hermite_polynomial_h; 23:38 updates backends.json.

Interpretation: FlagGems’ 11 commits cover a broad range but with a clear main line: the KMCompiler generation path expands its operator surface on the NVIDIA side (GRU/rank/dist operators) and adds tests, the FlagTune cost model moves toward multi-platform, and a W4A8 quantized linear operator appears on the CPU/ARM side — the latter corresponds to quantized inference needs for domestic server ARM (Kunpeng/Phytium-class) or edge scenarios, and together with FlagGems’ earlier support for the SiliconFlow backend (a cloud vendor) forms “cloud, edge, CPU” multi-form coverage. The Ascend multinomial fix and the Ascend operator fill-in in 1.4 are both part of deepening Ascend adaptation.

1.6 Other Components: FlagDNN Adds a Moore Threads Backend, vllm-plugin-FL Scheduling and Quantization Fixes, FlagScale Double Commit (09-04)

Source: FlagDNN (33958a1c25), FlagDNN (f619b92706), vllm-plugin-FL #439 (4b570bcd27), #440 (8174811506), #441 (396920a4cb), FlagScale #1281 (83617e0b66), FlagScale #1279 (09295660ba)

  • FlagDNN: 09-04 10:40 adds the mthreads (Moore Threads/MUSA) backend — following the addition of the hygon CPU reference implementation on 09-02, the DNN operator library gains another vendor backend; 13:54 fixes Ascend 910b4-4 adaptation (an adaptation correction for the Ascend 910B4 variant).
  • vllm-plugin-FL: three consecutive commits on 09-04 15:07~15:08 — #439 adds support in the scheduling layer for appending FlagGems blacklist entries (users/deployers can exclude specific operators from FlagGems dispatch, falling back to the reference implementation); #440 maintains registration of the available reference fallback (dispatch-layer fault tolerance); #441 has the quantization layer inherit native MXFP8 candidates for OOT (out-of-tree) quantization schemes — the “fallback-able, inheritable” semantics of both the dispatch and quantization paths are completed.
  • FlagScale: two commits on 09-04 19:58 — #1281 adds profiler kernel reports (kernel-level reports in profiling output); #1279 fixes the integration of engram with transformer (engram is an internal FlagScale module, fixing the correctness of its integration with transformer).

Interpretation: The long-tail components continue the theme of “vendor backends + fallback semantics”: FlagDNN follows up its CPU (hygon) support with a Moore Threads GPU backend, and together with the Ascend 910b4-4 correction, it now covers three platforms; vllm-plugin-FL’s three changes — blacklist/fallback/MXFP8 inheritance — all point to one goal: making the “FlagGems takeover + reference implementation fallback” dispatch model more controllable in real deployments (the blacklist allows FlagGems to be disabled operator by operator, preventing an individual misbehaving operator from blocking the entire inference chain). Such mechanisms are essential engineering for a plugin to enter production environments.

II. News Coverage and Ecosystem

2.1 Fourth Consecutive Quiet Window at the Component Level, Release Plateau Continues (09-04~09-05)

Source: 21 Google News RSS queries in Chinese and English (see appendix)

  • In this window (09-04 10:18 ~ 09-05 10:18), Chinese and English queries for FlagOS and its core components (FlagGems/FlagScale/FlagTree/FlagPerf/FlagAttention/FlagCX/KernelGen) once again returned zero hits across the board (direct component-level hits empty for the fourth consecutive window); “BAAI open source” and English FlagOS queries likewise produced no valid hits; the four HN Algolia hits (Gage tool/Lenovo tablet/Xfinity domain/VW layoffs) were all irrelevant noise and were excluded.
  • Trend over the past 3 days (corroborated by GitHub): After the dense 2.2 RC0 verification tag period (08-31~09-01) ended, 9/2~9/5 entered a “release plateau + multi-vendor delivery matrix construction period”: no new tags, no new releases, but build-infra completed Iluvatar CoreX/Cambricon vLLM 0.24.0 verification (1.2), sglang 0.5.18 vendor application environment landed (1.3), and FlagTree integrated the Qingwei backend (1.1). The next public action for 2.2 remains either rc0.postN iteration or cutting the official rc; the annotated dates of verification-period tags such as vllm-plugin-FL v0.3.0-rc0.post1 were verified as 08-31 (UTC), not new additions within the window.
  • High-frequency hits on the Chinese and English news side remain concentrated on the financials of member organizations (Enflame IPO pricing/loss breakdown, Moore Threads half-year report and new products, the “quality of turnaround” dispute between MetaX and Iluvatar CoreX, Cambricon software talent recruitment — see Chapter III). The ecosystem narrative is in a “peak capitalization period + earnings digestion period,” with no new official release from FlagOS itself to carry it forward.

Source: Sina Finance (Modu OPC report), Sina Finance original page (智源新境 and Ipsos China)

  • On 09-04, a Chinese query returned a hit for “智源新境 and Ipsos China reach strategic cooperation” (co-building a brand research paradigm for the AI era). Upon verification: 智源新境 (NextBrandAI) is a vertical marketing AI technology company registered in Shanghai Modu Space, unrelated to the Beijing Academy of Artificial Intelligence (BAAI) — purely a name collision, excluded, and not recorded as a BAAI institutional development (same conclusion as yesterday’s verification of iPolloWork OPC).
  • BAAI community content remains primarily third-party content aggregation (reposts on GPT-6 release teasers, Meta open-source models, Fei-Fei Li’s Atlas world model, etc.), not BAAI’s own research output; there were no new BAAI official model/platform/FlagOS-related announcements within the window.

III. Deep Dive into Member Units

3.1 Enflame: IPO Pricing Set at RMB 142.18 / Market Cap ~RMB 61.2B, H1 Loss Widens to RMB 632M, Roadshow Gives 2026/2027 Profitability Timeline (09-04)

Sources: Sina Tech / Operator Finance, “Enflame Technology’s H1 Loss Widens to RMB 632M”, Phoenix Finance, “Self-Developed Non-CUDA, Inference Migration Cost Only 1 Person-Day, So Why Still Losing Money?”, Jiemian News, “Tencent Becomes Biggest Winner of Enflame Technology’s IPO”, Sina Finance / Guandian, “Tencent Strategic Placement in Enflame Contributes RMB 248M”

  • Pricing and Subscription Wrap-Up: Enflame Technology (688801.SH) finalized its issue price at RMB 142.18/share, implying a market cap of approximately RMB 61.187B at listing; procedural news such as allotment numbers and institutional placements appeared in concentrated coverage on 09-04 (China AMC / Tianhong Fund received allocations; four bank wealth management subsidiaries collectively received over 23M shares), with Tencent’s strategic placement contributing RMB 248M, corresponding to a shareholding value of approximately RMB 12.6B. Jiemian News, in a piece titled “Tencent Becomes the Biggest Winner,” estimated its stake value at over RMB 10B — the dual identity of “largest customer + largest external shareholder” pervaded the entire IPO narrative.
  • Loss Data and Profitability Timeline: Media cited the prospectus/roadshow figures — 2026 H1 revenue of RMB 1.12B (already exceeding full-year 2025’s RMB 990M, +279% YoY), net loss attributable to parent of RMB 632M, widening by 3.57% YoY; cumulative losses over 2023–2025 exceeded RMB 4.3B (R&D expense ratio peaked at 408%); Chairman Zhao Lidong stated during the IPO online roadshow that he expects consolidated profitability in 2026 or 2027, and provided guidance of RMB 2.3–3.0B revenue for Jan–Sep 2026 (+325.78%–455.36% YoY).
  • Software Strategy Draws Renewed Attention: Phoenix Finance on 09-04 revisited the topic under the headline “Self-Developed Non-CUDA, Inference Migration Cost Only 1 Person-Day, So Why Still Losing Money?” — Enflame takes a domain-specific architecture (DSA) approach, develops its own “TopsRider” software platform, and is not CUDA-compatible (per Guancha); officially, inference scenario migration cost can be as low as 1 person-day. However, customer concentration — Tencent contributed over 80% of 2025 revenue (83.79% per Operator Finance) — and persistent losses remain the market’s focal concerns.

Analysis: Enflame’s IPO has entered the “pricing finalized, awaiting listing” stage (issue market cap ~RMB 61.2B, the last of the four “little dragons” to list on A-shares). Public attention is shifting from subscription gains to two substantive questions: whether the software ecosystem narrative can deliver (self-developed non-CUDA stack + 1 person-day inference migration cost — its enflame backend has already completed the vLLM 0.24.0 delivery loop and sglang 0.5.18 environment deployment in FlagOS build-infra, with the open-source unified stack serving as third-party corroboration of its “reduced migration cost” claim) and customer/shareholder concentration risk (over 80% of revenue from Tencent). The time window between the 3.57% H1 loss widening and its “2026/2027 profitability” expectation will be the checkpoint for every quarterly report post-listing.

3.2 Moore Threads: “Lushan” GPU to Launch This Year, First Domestic Hardware Ray Tracing; H1 Revenue +147% with Significantly Narrowed Loss; PD Heterogeneous Partnership with Qujing Technology (09-04/09-05)

Sources: Sina Finance / Kuaikeji, “‘Lushan’ GPU Debuts Hardware Ray Tracing”, Drivers Home / Kuaikeji Original, Sina Finance, “Revenue Up 147% Yet Still Loss-Making”, East Money, “Qujing Technology and Moore Threads Forge Strategic Partnership, Domestic PD Heterogeneous Solution Enters Production Environment”, PingWest (Partnership Report)

  • “Lushan” GPU (Consumer Graphics Flagship): At the 2026 H1 earnings briefing on the afternoon of 09-03, Moore Threads confirmed that “Lushan,” its high-performance graphics rendering chip based on the fifth-generation “Huagang” architecture, is expected to launch by end of 2026 (Q4). It is the first domestic graphics card to support hardware real-time ray tracing (supporting DXR/DirectX 12 Ultimate, integrating a second-generation hardware ray tracing engine and first-generation AI generative rendering architecture MTAGR 1.0); official figures versus the previous-generation MTT S80: 15x improvement in 3A game rendering performance, 50x improvement in ray tracing performance, 64x improvement in AI compute performance, with up to 64GB VRAM (benchmarked against RTX 50 series). FlagOS-relevant point: Lushan features a unified task engine UNITE, and the MUSA software stack already covers mainstream inference frameworks including vLLM, SGLang, and Ollama — local large-model inference is part of its product narrative.
  • Semi-Annual Report Figures: 2026 H1 revenue of RMB 1.736B, +147.42% YoY (already exceeding full-year 2025), R&D investment of RMB 769M (+38.16%); net loss attributable to parent narrowed to approximately RMB 12M (media figure: -RMB 11.5631M, versus -RMB 271M in the same period last year), approaching break-even; MUSA ecosystem developers surpassed 800,000 (reported yesterday). Multiple media outlets offered cautious interpretations under headlines such as “Revenue Surges Yet Still Loss-Making: Can Kuae + New Chips Support Profitability Expectations?”
  • Qujing Technology Strategic Partnership: From the evening of 09-04 (PingWest/Sina/East Money, multiple sources), reports emerged that Qujing Technology and Moore Threads have formed a strategic partnership to “produce high-quality compute with domestic cards,” with a domestic PD (Prefill-Decode disaggregated) heterogeneous solution entering the production environment — the combination of PD-disaggregated inference architecture with Moore Threads GPUs points toward cost-effective domestic AI token supply.

Analysis: Moore Threads presented a “earnings briefing + new product + new partnership” trifecta within this window: if the Lushan GPU launches on schedule in Q4, it will mark the domestic graphics stack’s first landing at the hardware ray tracing level, and its positioning of “graphics + AI + inference framework coverage” resonates with FlagOS’s inference-side adaptation (mthreads backend delivered 19 operators in half a day yesterday; today FlagDNN adds mthreads backend + FlagTree mthreads3.6 fixes) — the MUSA software stack’s stated coverage of vLLM/SGLang and the dense integration of Moore Threads on the flagos-ai side mutually corroborate each other. The Qujing Technology PD heterogeneous partnership is an industry signal of inference architecture (disaggregated inference) combining with domestic cards, representing one end of the same technological trend as the PD disaggregation support for FlagCX KV transfer in sglang-plugin-FL (merged 09-01).

3.3 MetaX: Book Profit of RMB 612M Turns Positive on RMB 887M in Unrealized Gains, Non-GAAP Still Loss-Making at RMB 48.86M, Mainstream Media Dissect “Quality of Turnaround” (09-04)

Sources: Sina Tech, “Behind MetaX’s Turnaround: ‘Recovery’ Isn’t from Selling Chips”, Sina Finance, “RMB 260B Market Cap MetaX Delivers First Semi-Annual Report”, Baike TA Shuo, “MetaX Semi-Annual Report: The ‘Financial Illusion’ Behind It?”

  • On 09-04, media concentrated on dissecting MetaX (688802.SH)’s first semi-annual report post-listing: H1 revenue of RMB 1.324B (+44.67%), net profit attributable to parent of RMB 612M, book turnaround to profitability (versus a loss of RMB 186M in the same period last year); however, non-GAAP net profit remained at -RMB 48.86M (loss narrowed 75.83% YoY).
  • Breakdown of Turnaround Sources (consistent across multiple sources): The approximately RMB 660M gap between book profit and non-GAAP figure comes from RMB 887M in fair value change gains (105.75% of total profit) — the company used idle raised funds for cash management and external investments, with the bulk contributed by unrealized gains on a listed company equity stake held through subsidiary “MetaX Digital Intelligence”; media questioned the quality under headlines such as “Recovery Isn’t from Selling Chips” and “Financial Illusion.”
  • Operational highlights and concerns coexist: consolidated gross margin of approximately 57.2% (+1.1pct), direct sales share jumping from 41.9% to 64.7% (+123.6%); R&D expenses of RMB 525M (39.65% of revenue), R&D headcount of 747 (70.4%); net operating cash flow of -RMB 1.297B (approximately RMB 400M more outflow YoY), prepayments of RMB 2.196B (+163.24%, strategic stockpiling to secure wafer foundry and raw material supply); receivables aged 1–2 years expanded 12.5x in half a year.

Analysis: MetaX became the first among the “four little dragons” to report book profitability, but the scissors gap between “RMB 612M net profit vs. non-GAAP -RMB 48.86M” led media consensus to conclude “insufficient quality of turnaround, core business still loss-making.” From a FlagOS perspective: MetaX’s ecosystem-side investment has not been affected by financial fluctuations (yesterday both vLLM/sglang images for metax-maca3.8.1.3 were registered), while cash flow pressure (-RMB 1.3B operating cash flow + RMB 2.196B prepayments) and the approaching September lock-up expiry are variables more worth tracking — its surging direct sales share implies volume orders from intelligent computing centers/telecom carrier key accounts, directly relevant to the FlagOS unified stack’s “adapt once, deliver across multiple chips” scale-up requirement scenario.

3.4 Iluvatar CoreX: Book Turnaround of RMB 106M Entirely Supported by SJ Semiconductor Unrealized Gains, Gross Margin Plummets to 17.2%, Core Business Loss Widens (09-04)

Sources: Times Finance / Sina, “Behind Iluvatar CoreX’s Turnaround: Revenue Up 191%, Over RMB 700M in Equity Investment Unrealized Gains Underpin Profit”, Sohu (09-04 Reprint)

  • Iluvatar CoreX (09903.HK)’s first interim results post-listing continue to be digested: H1 revenue of approximately RMB 946M (+191.6%, already 91.4% of full-year 2025), book net profit of RMB 106M achieving turnaround; however, profitability came almost entirely from RMB 756M in fair value unrealized gains on its stake in SJ Semiconductor (688820.SH, packaging & testing, listed on STAR Market in April) — unrealized, no cash inflow, with a 12-month lock-up period — excluding which the core business actually lost approximately RMB 654M, widening from RMB 609M in the same period last year.
  • Business structure: General-purpose GPU contributed RMB 916M (96.9%), with inference chip “ZhiKai” series revenue surging 651.8% YoY; however, consolidated gross margin plummeted from 50.1% in 2025 H1 to 17.2%. Financing: cumulative fundraising exceeding HKD 10.7B in 2026 (January IPO + HKD 10B-level H-share placement + capitalization issuance plan), multiple rounds of “cash infusion” within 8 months of listing.

Analysis: Iluvatar CoreX (FlagOS member, Iluvatar backend)’s “turnaround” is structurally identical to MetaX’s — both rely on equity investment unrealized gains rather than GPU core business (yesterday’s “Four Little Dragons Interim Report Perspective” already noted the commonality of two book turnarounds; this window provides itemized breakdowns from media). Two points worth noting from a FlagOS perspective: ZhiKai inference chip revenue +651.8% confirms that inference volume ramp is the real mainline (resonating with build-infra records showing its vLLM 0.24.0 F-path validation passed while T-path is blocked by toolchain defects — inference chip ramp requires the software stack to keep pace); gross margin collapsing from 50.1% to 17.2% signals price war and cost pressure, which will affect judgment on the sustainability of its ecosystem/software investment.

3.5 Cambricon: 11 Teams Aggressively Recruiting Software Talent, Doubling Down on AI Software Ecosystem (Posted 09-02 / Reported 09-04)

Sources: Lianhe Zaobao, “Cambricon Aggressively Recruits Software Talent, Doubles Down on AI Software Ecosystem”, Guandian (Recruitment Announcement Report), Sina Finance / China Fund News, “Cambricon Makes a Big Move”

  • On 09-02, Cambricon published a recruitment announcement on its official WeChat account, planning to recruit software talent for 11 teams: driver team, compiler team, high-performance communication team, compute library team, deep learning framework and ecosystem component development team, inference and training solution team, etc., with the goal of “building a leading AI software system.”
  • The announcement’s logic (interpreted by Lianhe Zaobao on 09-04 as “emulating Nvidia’s CUDA software ecosystem playbook”): the proliferation of AI code generation tools is reshaping software development paradigms, and the rising importance of post-training is dramatically shortening model iteration cycles, making the market more in need of “ultra-large-scale, well-ecosystem-compatible, highly stable” AI infrastructure — the software stack becomes the core competitive moat.
  • Semi-annual report figures corroborate: Cambricon’s 2026 H1 continued optimization of its foundational system software platform — on the training side, focusing on ecosystem building, model adaptation and performance optimization, and low-precision training; on the inference side, focusing on large-model inference, computation graph compilation, operator programming, acceleration libraries, and communication.

Analysis: Cambricon (FlagOS member, cambricon backend) has elevated “software ecosystem” to a strategic height on par with hardware, consistent with its active pace on the FlagOS side — in this window, build-infra just completed validation of its vLLM 0.24.0 application image (#718) and deployed the application environment for sglang 0.5.18 (#730). The 11 recruitment team directions (compiler/compute library/communication/framework/inference-training solutions) map almost one-to-one onto the FlagOS component map. The dual-track approach of proprietary software stack (self-developed) + open-source unified stack (FlagOS co-building) is the standard form of current software investment among leading domestic chip vendors; Cambricon, as the largest domestic AI chip company by market cap, doubling down on software is a long-term positive signal for the FlagOS ecosystem.

IV. Summary

Main threads of the FlagOS ecosystem in this window (09-04 10:18 ~ 09-05 10:18):

  1. FlagTree integrates the Tsingmicro backend for the first time, adding another seat to the compiler backend matrix: #1071 introduces the TLE-DSA dataflow programming abstraction and the DsaToCore conversion at a scale of 300 files (aligned with Triton 3.6) — Tsingmicro becomes the latest member unit to enter the unified compiler, and it preserves the native expression of its reconfigurable architecture through an independent DSA dialect rather than simply emulating many-core.
  2. The vLLM 0.24.0 delivery matrix expands to Cambricon and Iluvatar, with the first record of F/T path divergence: cambricon and iluvatar-corex4.4.0 pass validation (#718), but the Iluvatar T path (vendor Triton 3.1.0) is judged undeliverable due to a uint32/int32 mixed-division defect in the apply_top_k_top_p kernel (#722), with the fix direction locked to forcing a pytorch sampler fallback at the plugin layer — “why it cannot be delivered” is precisely recorded, raising the credibility of delivery engineering.
  3. Construction of the sglang 0.5.18 vendor image matrix begins: application environments (deps_app/app env) for Cambricon and Enflame land (#729/#730), the sglang image tag for Enflame tops1.10.6 is registered (#734), and 12 build/validation fixes focus on wheel build imageization (rsproxy), network deduplication, and readiness gating — the sglang line is replicating the validation-delivery closed loop of the vLLM line.
  4. The Ascend toolchain completes the CANN 8.5 migration away across the full FlagOS stack: on the FlagGems-vllm side, cann850 is removed (#740, following yesterday’s FlagGems #5970), while 4 Ascend operators are added in batch (including the causal_conv1d family for Mamba-class models); inference hot-path optimizations such as Hygon fused QKV RMSNorm and Moore Threads fused add rms norm advance in parallel.
  5. Member units collectively enter a “capitalization peak + earnings digestion” period: Enflame’s issue price is set at 142.18 yuan with a market value of about 61.2 billion, H1 losses widen to 632 million but a profitability timetable for 2026/2027 and a revenue forecast of 2.3~3.0 billion for the first three quarters are given; Moore Threads’ “Lushan” GPU will launch within the year, debuting domestic hardware ray tracing and forming a PD heterogeneous partnership with Qujing Technology; both MetaX and Iluvatar are dissected as “turning profitable on paper while still loss-making on a non-recurring basis” (MetaX non-recurring -48.86 million, Iluvatar’s main business losses widening), and Cambricon’s team 11 is aggressively recruiting software talent — each company’s software investment on the FlagOS side is heating up in sync with its capital narrative.
  6. The quiet window of zero component-level news hits extends to a fourth one; the high activity of 51 commits on the GitHub side shows that engineering progress continues to diverge from external visibility — FlagOS remains in a “heads-down construction” phase, and the next official public action is expected to still be the 2.2 rc iteration (signaled by activity in the FlagRelease repository).

Appendix: Complete Source List

Source Verification Result
GitHub org repos API (52 repos) 11 repos pushed within the window (build-infra/FlagGems/FlagTree/FlagGems-vllm/vllm-plugin-FL/FlagDNN/FlagScale/sglang-plugin-FL/flagtree-cpu/docs/release-info); pushed_at changes for docs/release-info/flagtree-cpu/sglang-plugin-FL were verified via default-branch commits as non-default-branch activity (0 commits on the default branch within the window), and are not counted in the main text; no new repos created
GitHub commit search All 51 commits within the window retrieved (build-infra 19, FlagGems 11, FlagGems-vllm 9, FlagTree 5, vllm-plugin-FL 3, FlagDNN 2, FlagScale 2); note: committer.date returned by commit search mixes two timezone formats, Z and +08:00; each entry was re-verified after converting to Beijing time according to its respective timezone
GitHub tags/releases API + annotated tag object verification No new tags, no new releases within the window; the two 2.2 RC0 validation-period tags — vllm-plugin-FL v0.3.0-rc0.post1 (08-31 18:03Z) and FlagGems-vllm v0.2.0-rc0.post1 (08-31 18:02Z) — were both created before the window; the latest releases of active repos are all from 06-24 (FlagOS 2.1) or earlier
FlagTree #1071 300 files, first substantive Tsingmicro integration (of 160 tsingmicro commits in the org, only this 1 was substantively merged into FlagTree; the other 2 were docs/unrelated); TLE DSA dialect + DsaToCore + tsingmicro integration tests fully verified
FlagTree #1096/#1101/#840/#899 Iluvatar llvm22 v0.6.1 and delivery process updates, Iluvatar non-ivcore11 assembly skip, TLE reshape/named barrier memalloc/gpu.copy mask/barrier arrival, TLE remote node2 — verified item by item
build-infra #718-#736 (19 items) Verified item by item: cambricon+iluvatar vLLM 0.24.0 validation passed (#718); iluvatar T path not deliverable — corex Triton 3.1.0 rejects uint32/int32 mixed division, sampler patch coverage gap (#722, full commit body verified); cambricon 0.20.2 re-validation (#724); sglang 0.5.18 application environment landed for enflame/cambricon (#729/#730); shim full-index release (#719); build fixes #720/#721/#723/#725/#731/#732/#733/#735/#736; status matrix and validation records #726/#727/#728
FlagGems-vllm #731/#728/#729/#730/#740/#738/#735/#697/#739 Ascend 4 operators (add_rms_norm/swiglu/causal_conv1d_update/causal_conv1d_fn) verified item by item; cann850 removal; Hygon fused QKV RMSNorm optimization; mthreads fused add rms norm; fp8/fp4 MQA logits TLE; circular import fix
FlagGems 11 items KMCompiler NVIDIA gru (#6002)/matrix rank (#5942)/dist operator refinement (#5958); FlagTune multi-platform Mul cost model (#5762); ARM CPU W4A8-G128 linear (#5904) and setup.sh aarch64 fix (#5999); Ascend multinomial fix (#5980); slice/polynomial fixes; backends.json update
FlagDNN 2 items New mthreads backend (33958a1c25), fix for Ascend 910b4-4 adaptation (f619b92706)
vllm-plugin-FL #439-#441 FlagGems blacklist addition, reference fallback retained, OOT quantization inherits native MXFP8 candidates
FlagScale #1281/#1279 profiler kernel reports; engram transformer integration fix
Google News RSS Chinese set (14 groups) Zero hits for FlagOS/component-level in 24h (fourth consecutive window); hits were all financial-side coverage of member units: Enflame (pricing 142.18 yuan/61.187 billion market cap, allotment number 20657 re-circulated, H1 loss 632 million, roadshow profit expectations, Tencent strategic placement 248 million/market cap 12.6 billion, funds and wealth management subsidiaries allotted), Moore Threads (Lushan GPU hardware ray tracing, revenue +147%, Qujing Technology PD partnership), MetaX (book turnaround 612 million vs. non-recurring -48.86 million, unrealized gain breakdown), Iluvatar CoreX (unrealized gain turnaround, gross margin 17.2%, Zhikai +651.8%), Cambricon (11 teams software recruitment); “Zhiyuan Xinjing and Ipsos partnership” was verified via the original site as a Shanghai marketing technology company that collides with BAAI’s name, and was excluded
Google News RSS English set (7 groups) No valid hits for FlagOS/BAAI/component English queries (1 Enflame IPO-related item was a reprint of an English financial roundup); all 4 HN Algolia hits were irrelevant noise and were excluded
Financial media original-site verification MetaX (Sina Tech 09-01 in-depth + Baike TA Shuo breakdown), Iluvatar CoreX (Times Finance 08-31 + Eastmoney 09-01), Enflame (Sina Tech/Operator Finance 09-04), Moore Threads Lushan GPU (Kuaikeji/ITHome 09-04 earnings call report), Cambricon (Lianhe Zaobao 09-04/Guandian/China Fund News 09-02) — each verified by title + body summary
BAAI Community (gnews hits) All third-party content aggregation (GPT-6 teaser, Meta open-source models, Fei-Fei Li Atlas, etc.), not BAAI’s own achievement releases; no official BAAI announcements within the window