Research window: 2026-10-30 06:00 ~ 2026-11-02 06:00 (Beijing time, 72 hours; Monday morning report, covering Saturday and Sunday, window continues from the previous daily report’s cutoff).

Sources: GitHub (full audit of 61 repos under org: flagos-ai, 256 commit searches, 115 merged PRs, 275 PRs under review, 100 org event stream entries, atomic audit of Releases in 31 repos and Tags in 8 repos, atomic re-audit of 5 repos with branch attribution verification, verification of key PR bodies, audit of community release list), Google News RSS (24 sets of Chinese and English query terms, via proxy), HN Algolia (8 sets), CSDN FlagOS column (per-article timestamp verification), BAAI Community, official websites and financial report coverage of nine member organizations (see appendix source list for details).

Index

  • Today’s Highlights: Ascend 910C native quantum operator line connected end-to-end (167 hardware acceptance items); 2.3.0 two image-layer changelogs in place; quantum synthesis compiler 35 merges in a single window (10-30 ~ 11-02)
  • I. Open Source Project Progress (GitHub Activity)
    • 1.1 Org cadence: 17 repos pushed, 256 commits, 115 merges, 275 under review (10-30 ~ 11-02)
    • 1.2 Version landscape: 2.3.0 image layer wrapping up, Hygon DTK 26.10 and AMD ROCm 7.2.4 both in place (10-31 ~ 11-01)
    • 1.3 FlagQuantum: Ascend 910C native support merged, 167 hardware acceptance items passed (10-31)
    • 1.4 FlagQuantum: 35 merges in a single window, synthesis-line resource accounting enters contractualization phase (10-31 ~ 11-02)
    • 1.5 FlagGems: Three KernelGen operators enter NVIDIA line, Kunlunxin XPU3 ten performance patches closed out (10-30)
    • 1.6 Torch-FL: Horizon BPU passes convolution and pooling, Tianshu, MetaX, Hygon paths converge separately (10-30 ~ 11-01)
    • 1.7 FlagCX: Hygon bitcode and Device IR unified packaging, SymMem coverage tightened per adapter (10-31 ~ 11-01)
    • 1.8 Other repos: FlagPrism integrates MetaX, optest integrates FlagTrain, FlagTree reports to Feishu (10-30 ~ 11-02)
  • II. News Coverage and Ecosystem
    • 2.1 Channel audit: No new hits in CSDN column or HN within window, zero component-level news items (11-02)
    • 2.2 Member organizations’ capital markets: Enflame and MetaX Q3 reports disclosed same day (10-30)
    • 2.3 Ecosystem watch: World Internet Conference Wuzhen Summit opens today, theme focuses on open source and openness (11-02)
  • III. Deep Dive into Member Organizations and Backend Vendors
    • 3.1 Hygon: DTK 26.10 21 merges from base image to wheel (10-31 ~ 11-01)
    • 3.2 Ascend: Beyond 910C native quantum line, XCS interface and communication CI advance in parallel (10-30 ~ 11-01)
    • 3.3 Kunlunxin: XPU3 ten performance patches merged in a single day (10-30)
    • 3.4 MetaX: FlagPrism adds MACA backend, attention side adds Parallax and Hy3 (10-30 ~ 11-02)
    • 3.5 Tsingmicro, Enflame, Tianshu: TXDA enters vLLM 0.28, GCU graph mode, FlagTree version pinning (10-30 ~ 11-01)
    • 3.6 Moore Threads and Horizon: MUSA side adds large-model attention, BPU side adds operator paths (10-30 ~ 11-01)
  • IV. Summary and Trend Observations
  • Appendix: Source verification table, complete source list

Today’s Focus: Ascend 910C native quantum operator path fully connected (167 hardware acceptance checks); changelogs for two 2.3.0 image layers in place; 35 merges into the quantum synthesis compiler in a single window (10-30 ~ 11-02)

Date: 2026-10-30 ~ 2026-11-02　Sources: FlagQuantum PR #958 · FlagQuantum PR #970 · build-infra PR #1584 · FlagQuantum PR #1000 (full list in appendix)

  • Quantum: FlagQuantum #958 moves Ascend 910C from the “experimental split-FP32 path” to the “native provider operator” path and merges it into mainline. The final merge hardware record given in the PR body is 167 passed in 115.41s — 142 circuit, noise, TEBD, and QEC cases, plus 25 native dynamics cases, including two cold-process plan replays. The stack is pinned to Python 3.11.15, PyTorch 2.10.0+cpu, CANN 9.0.0, FlagTree 0.7.0+ascend3.5, FlagGems 5.4.0, and FlagCX v0.14.1 (with pure Ascend/FlagOS patches, not linking torch_npu). On the same day, #970 adds a fail-closed native 910C regression lane on the pre-provisioned ARM64 runner, and the CI side has several additional pins for wheel provenance and action archiving.
  • Release: All 21 merges into build-infra over the three days point to 2.3.0 — a new hygon-dtk26.10 image layer (base image switched to ROCm-aware Open MPI, runtime adds flagcx==0.14.1+dtk2610, removes the ineffective HIP_VISIBLE_DEVICES=all), while the AMD rocm7.2.4 layer gets its flagcx pin and amdsmi dependency completed, and the changelog date and image description for 2.3.0 are recorded layer by layer. However, the release/2.3 manifest directory in the community repo has still not been created, and no 2.3.0 tag has appeared on the release-info site, indicating this round is pre-release image-layer preparation, not the release action itself.
  • Compiler: Of the 35 merges into FlagQuantum in this window, 28 land on the synthesis line, wiring Qmod public input admission, register borrowing, Const calls, unsigned comparison, post-computation cleanup and zero-slot reuse, within/apply scoping, and bounded exponentiation into the existing synthesis pipeline item by item; #1000 gives a recomputable example of the change in the resource ledger — x * 256 drops from a baseline of 28 allocated qubits and a 4 GiB state buffer to 19 allocated qubits and 8 MiB, with 12 operands and depth 4.
  • Assessment: The three lines point to the same thing — the form of evidence for quantum-stack multi-chip adaptation is shifting from “it runs on some chip” to “a reproducible hardware acceptance record + a verifiable resource ledger + a CI lane that goes red on failure.” For 2.3.0, the image layers are already complete; what remains is the two release actions: manifest freeze and site go-live.

I. Open-Source Project Progress (GitHub Activity)

1.1 Organizational Cadence: 17 Repos Pushed, 256 Commits, 115 Merged, 275 Under Review (10-30 ~ 11-02)

Date: 2026-10-30 ~ 2026-11-02　Source: flagos-ai org repositories · Commit search · Merged PR search · Open PR search

  • Repo surface: 17 of 61 repos had pushes within the window (previous period: 21 repos on a 24-hour basis). Ordered by push recency: docs, FlagQAI, FlagQuantum, FlagPrism, Torch-FL, FlagTree, release-info, build-infra, vllm-plugin-FL, FlagCX, optest, FlagGems-sglang, FlagGems, Megatron-LM-FL, FlagGems-vllm, OpenCourse, FlagSparse. The docs push landed on a documentation PR branch, release-info on gh-pages, and vllm-plugin-FL on a PR branch; none of the three had new commits on their default branches within the window.
  • Commit surface: Commit search returned 256 hits (previous period: 189 on a 24-hour basis), split as FlagQuantum 148, FlagQAI 42, build-infra 21, FlagGems 16, Torch-FL 13, optest 5, FlagCX 3, OpenCourse 3, FlagTree 2, and 1 each for FlagSparse, FlagGems-vllm, and FlagPrism. That is roughly 85 per day, below the previous period’s 189 in a single day, with a clear weekend effect; FlagQuantum and FlagQAI alone account for 190 commits, or 74%.
  • Merge surface: 115 merged PRs (previous period: 183 on a 24-hour basis) from 14 repos: FlagQuantum 35, build-infra 21, FlagGems 18, Torch-FL 13, FlagQAI 11, optest 5, FlagCX 4, FlagTree 2, and 1 each for FlagPrism, FlagGems-sglang, Megatron-LM-FL, FlagGems-vllm, OpenCourse, and FlagSparse. That is about 38 per day, roughly one-fifth of the previous period’s 183 in a single day.
  • Review surface: 275 open PRs were updated within the window (previous period: 253), with 146 newly opened (previous period: 84). By repo: FlagGems 158 (82 new), FlagGems-sglang 18 (18), FlagTree 15 (5), FlagGems-vllm 15 (8), Torch-FL 13 (11), vllm-plugin-FL 13 (7), FlagTrain 9 (1), FlagAttention 8 (5), sglang-plugin-FL 7 (1), docs 6 (3), FlagScale 5 (1), FlagCX 3 (1), and 1 each for FlagQAI, FlagQuantum, flir, FlagBLAS, and whisper-triton.
  • Events and branches: The org event stream had 100 entries: 37 pushes, 13 branch deletions, 23 PR events, 11 issue comments, 10 branch creations, 2 PR reviews, 3 review comments, and 1 fork. Most activity concentrated in FlagQuantum, build-infra, and FlagQAI. For the 5 pushed repos not covered by commit search, an atomic cross-check found: Megatron-LM-FL merged #203 at 10-30 16:55 (within the window); the 10-31 FlagGems-sglang push was on branch feat/qwen36-remaining-native-validation, with #251 merged to main the same day; the 11-01 vllm-plugin-FL push was on a PR branch, and the default branch’s latest commit remains GLM-5.3 support from 10-30 00:38, which belongs to the previous window; the docs push was on a documentation PR branch (default branch’s latest commit is 10-03); both release-info pushes were on gh-pages, an automatic refresh of the release site.
  • Comparison: Versus the previous 24-hour window, merge volume fell about 80% on a daily-average basis and commit volume about 50%, yet total open PRs rose another 8.7% and new opens nearly doubled. This is the classic “low weekend output, growing backlog” pattern, with the backlog driven mainly by FlagGems’ batch of KernelGen filings (82 new).

1.2 Release Surface: 2.3.0 Image Layers Wrapped Up, Hygon DTK 26.10 and AMD ROCm 7.2.4 Both in Place (10-31 ~ 11-01)

Date: 2026-10-31 ~ 2026-11-01　Source: build-infra PR #1584 · build-infra PR #1568 · build-infra PR #1571 · build-infra PR #1586 · release-info site · community 2.2 release manifest

  • Image layer one (Hygon): The hygon-dtk26.10 layer was built from scratch — #1568 installs ROCm-aware Open MPI in the base image, #1577 publishes that layer’s flagcx wheel with a +dtk2610 suffix, #1578 bootstraps the runtime image without pinning flagcx, and #1584 consolidates the pin and its changelog into a single PR. The same layer also saw three bot changelog date records (#1575, #1579, #1585) and #1587, which hands off the defect list for the Hygon-side vendor Triton 3.7.0 to the vLLM line.
  • Image layer two (AMD): #1571 pins flagcx into the rocm7.2.4 runtime dependencies, #1576 removes the ineffective HIP_VISIBLE_DEVICES=all environment variable, #1586 adds amdsmi, and #1572, #1580, #1582, and #1588 record changelog dates. The base image layer additionally saw #1569, #1570, and #1583 recording 2.3.0 dates and refreshing the image description.
  • Release-state check: An atomic check of Releases across 31 repos and Tags across 8 repos found no new Release objects within the window; the only timestamp update was the FlagCX v0.14.1 tag (refreshed 11-01 00:10), while the latest official version on the GitHub Releases API remains v0.14.0, released 2026-09-28 (its body states it corresponds to FlagOS 2.2). The community repo’s release/2.3/release-2.3.yaml returns 404, and the release/2.2 manifest is still present, indicating 2.3 has not yet entered the manifest-freeze stage.
  • Watch items: Of the 21 merges, 8 are bot changelogs, 2 are description refreshes, 6 are dependency and bootstrap fixes, and 5 are CI and docs. Following FlagOS 2.2’s release cadence (manifest freeze → image tagging → release-info go-live), 2.3.0 still lacks the manifest and site steps; if the next work week continues with only in-layer patching and no manifest freeze, the release window will be compressed.

1.3 FlagQuantum: Native Ascend 910C Support Merged, 167 Hardware Acceptance Tests Pass (10-31)

Date: 2026-10-31　Source: FlagQuantum PR #958 · FlagQuantum PR #970 · FlagQuantum commit c417d80

  • Acceptance data: The PR body records the combined hardware run result as 167 passed, 293 warnings in 115.41s; two integration files cover 142 circuit, noise, TEBD, and QEC cases plus 25 dynamics cases, including two cold-process plan replays. It also reports a CPU-side regression of changed paths at 1332 passed, 3 skipped in 21.04s, with 23 capability and evidence-field contract tests passing.
  • Coverage: complex64/complex128 state vectors, MPS, tensor networks, and density matrices; autograd and explicit gradient methods; hybrid training, noise, and TEBD; native Ising/Lindblad/Rydberg dynamics and instrumentation; QEC pre-decoder training; owner-sharded state-vector, MPS, and tensor-network training. The PR explicitly states that CPU defaults, public signatures, and schema versions are unchanged, and that what is added is the additive native dynamics capability already approved in the device capability proposal.
  • Dependency boundaries: Torch-FL is taken from commit 3aebd0c on branch feat/ascend-complex-lowering, and FlagCX from commit 6fe3631 of official v0.14.1 with pure Ascend/FlagOS patches; the PR states outright that “published wheels without these fixes are insufficient to reproduce,” effectively binding reproducibility to source commits rather than binary packages.
  • CI side: #970 builds a fail-closed native regression lane on a pre-provisioned ARM64 runner; companion commits archive-pin the actions used by native regression to verified commits, preserve native wheel provenance in the isolated runner image, override inherited wheel versions with a certified provider, and stream native test progress into the watchdog.
  • Significance: Previously the 910C had only evidence from an experimental split-FP32 path; this line pushes the quantum stack’s Ascend adaptation down to the native provider layer and writes hardware acceptance records into repo docs and CI rather than leaving them in personal environments.

1.4 FlagQuantum: 35 Merges in a Single Window, Synthesis-Line Resource Accounting Enters the Contract Stage (10-31 ~ 11-02)

Date: 2026-10-31 ~ 2026-11-02　Source: FlagQuantum PR #1000 · FlagQuantum PR #995 · FlagQuantum PR #993 · FlagQuantum PR #984

  • Volume and structure: FlagQuantum merged 35 PRs in 72 hours: 28 on the synthesis line, 1 for native Ascend support, 5 for CI and tooling, and 2 for simulation docs and fixes. The synthesis line is the clear majority, and all of it revolves around three things: submittable inputs, verifiable semantics, and accountable resources.

  • Input surface: #969 introduces a common Qmod input and shared Program semantics, #971 admits bounded Qmod files and text input, #985 admits native one-bit phase statements, and #993 and #994 support bounded within/apply scopes and their nesting. Qmod has gone from an internal format to an externally submittable input form.
  • Semantics surface: #972, #981, #982, and #990 successively add unsigned equality/less-than predicates, consistent value selection from Python conditions, unsigned relations and chained interval predicates, and native unsigned comparison expressions; #974, #984, #995, and #997 combine predicate phases with initialized-register borrowing, without relying on temporary flags.
  • Resource surface: #988 and #989 validate Const borrowing and fresh-output in register calls; #991 validates computation-result cleanup and storage reuse, #992 validates local ownership and zero-slot reuse; #999 cuts workspace for constant addition proven carry-free, and #1000 cuts temporary bits for exact powers-of-two multiplication. For x * 256, the baseline is 28 allocated qubits, 18 workspace bits, 14 operations, depth 6, and a 4 GiB state buffer; after the merge it is 19 allocated qubits, 9 workspace bits, 12 operations, depth 4, and an 8 MiB state buffer. The x * 4 probe drops from 14/12 allocated and used bits, 8 workspace bits, 24 operations, depth 18, and 256 KiB to 10/10, 4, 10, depth 4, and 16 KiB.
  • Boundaries and caveats: The PR states explicitly that these are “exact resource accounting,” not guarantees of runtime speedup, optimality, or peak process memory; zero, one, non-power-of-two constants, two quantum operands, and derived values with a zero lower bound still take the generic multiplication path; the change touches only one production file, the Compiler’s lowering.
  • Implication: For the past quarter the quantum synthesis compiler’s main thread was “how many constructs can be synthesized”; this window shifts to “can each construct’s resource accounting be recomputed and regression-tested.” The value of this work is not point performance but making resource budgets a regression-testable contract — a different kind of engineering from the mainline operator library’s performance patches.

1.5 FlagGems: Three KernelGen Operators Land on the NVIDIA Line, Ten Kunlunxin XPU3 Performance Patches Closed Out (10-30)

Date: 2026-10-30　Source: FlagGems PR #7988 · FlagGems PR #7957 · FlagGems PR #7672 · FlagGems PR #7257

  • KernelGen line: chx7514 merged three view-class operators in one batch — swapaxes_ (#7314), movedim (#7307), and swapdims_ (#7316), all Triton kernel implementations, going through the usual KernelGen auto-generation to human-review channel.
  • SiliconFlow line: Four from winfan1314: PPU triangular_solve default and out execution (#7672), PPU upsample_nearest2d_backward with explicit scales (#7668), PPU STFT via acFFT (#7667), and MetaX upsample_nearest2d_backward explicit scale mapping (#7589). All four align execution semantics for the same operator across two backends — cross-backend consistency work.
  • Kunlunxin XPU3 line: Ten in a single day, covering #7988’s optimization of 11 pointwise and reduction operators, #7957’s direct-pointer foreach kernels plus allclose/isclose reductions and the index_put/clone copy path, #8052’s BF16 pointwise 1D tiling port, #7960’s SPLIT_K KeyError in reusing mm tuning configs, #7983’s is_same_size and channel_shuffle performance, #8024’s exact fmod truncation (div_rn plus FMA residual), #7955’s single-launch neighborhood search for choose_qparams_optimized, #7954’s sgn bf16 per-dtype tiling, #8048’s libdevice stub symbol fix for foreach_pow with vendor powf fallback, and #8047’s test-routing fix for foreach_pow_scalar_.
  • Ascend line: #7257 switches the topk custom operator’s invocation to tle_raw.call, unifying it with the TLE dialect — an interface convergence, not a performance change.
  • Backlog: FlagGems has 158 open PRs and 82 newly opened, almost all from KernelGen batch submissions; the ratio of 18 merged to 158 open shows review capacity still has not caught up with generation, the most concrete risk before the 2.3 freeze.

1.6 Torch-FL: Horizon BPU Gets Convolution and Pooling Working; Iluvatar, MetaX, and Hygon Paths Converge Separately (10-30 ~ 11-01)

Date: 2026-10-30 ~ 2026-11-01　Source: Torch-FL PR #686 · Torch-FL PR #693 · Torch-FL PR #698 · Torch-FL PR #700

  • Horizon BPU: #686 routes conv, batch norm, and pooling from the eager path directly onto the BPU; #693 goes further, carrying avg_pool on the BPU via an equivalent depthwise convolution; #696 re-expresses avg_pool’s convolution form through the functional API. Taken together, the BPU side goes from “has operators” to “has a maintainable path.”
  • Iluvatar: #698 pins Iluvatar CI’s FlagTree to 0.7.0 on the grounds that this is the version actually running on the chip; #697 routes baddbmm through the vendor GEMM instead of the partial-tile FlagGems kernel. Both settle the question of which implementation is used.
  • MetaX and Hygon: #688 registers MetaX’s mm on the FlagGems side and relaxes the K dimension its Triton rejects; #701 gets Hygon DCU’s FlagGems SDPA tests running on lanes that can actually execute them. #701’s value is test credibility — a wrong skip reason masks failures long-term.
  • NVIDIA default path: #700 fixes the default NVIDIA conf routing 1,485 operators to the boxing kernel, the highest-impact change in this window; #684 converts FP4’s half-byte shift into tensor operations to hit the already-routed overload; #685 relaxes the contraction axis of soft-lowp GEMM per the platform dot lower bound; #687 narrows the day-0 FlagGems override to the operators its own description justifies; #689 falls back to the pinned index when the CPU-torch image fails; #692 keeps integration coverage at case granularity rather than file granularity.

1.7 FlagCX: Hygon Bitcode and Device IR Packaged Uniformly, SymMem Coverage Tightened by Adapter (10-31 ~ 11-01)

Date: 2026-10-31 ~ 2026-11-01　Source: FlagCX PR #667 · FlagCX PR #687 · FlagCX PR #688 · FlagCX PR #669

  • Hygon side: #667 is a structural change on the UIL line — it adds the Hygon bitcode build and unifies how Device IR is packaged. Previously each backend packaged its own IR, so Hygon onboarding required maintaining a separate set of artifacts; after unification, the packaging cost for new backends drops.
  • Test side: #687 enforces SymMem coverage in CI by adapter type, preventing “an adapter never ran symmetric-memory tests but the pipeline is still green”; #688 and #689 select the DU stream guard by torch generation and backport the same fix to 0.14.0-rc2.
  • Under review: #669’s Ascend 910C CI pipeline and unit-test orchestration, #686’s Ascend device completion tracking (proposed by the FlagQuantum side, serving the quantum stack’s communication-completion semantics), and #685’s DAG-based UniRunner collective-communication and fused-communication workflow. All three are open and constitute the pending merges for the 2.3 communication surface.

1.8 Other Repos: FlagPrism Onboards MetaX, optest Onboards FlagTrain, FlagTree Reports to Feishu (10-30 ~ 11-02)

Date: 2026-10-30 ~ 2026-11-02　Source: FlagPrism PR #18 · optest PR #19 · FlagTree PR #1457 · FlagQAI PR #15

  • FlagPrism: #18 merged at 11-02 00:11, completing the debugger runtime for the MetaX/MACA backend — a dedicated MacaRuntimeBackendAdapter handles device buffers, host staging, and stream-ordered transfers, with a new MACA.cmake build module, Python bindings, tests, and docs. This is the first time FlagPrism has brought MetaX into the debugging path.
  • optest: Five merges around test infrastructure — #19 uses uv to manage FlagTrain’s vendor dependencies and hooks into its CI, #20 fails fast on insufficient shared memory, #21 retries the package index on install failure, #22 raises S60’s shared memory, and #23 isolates concurrency by library and backend. The direction is separating “environment problems” from “operator problems” in logs.
  • FlagTree: #1457 writes FlagGems CI results into a Feishu multi-dimensional table, moving test results from repo pages to the team dashboard; #1494 temporarily disables Enflame’s 03-topk test, a stop-loss for a known failure.
  • FlagQAI: 11 merges and 42 commits, all in the web workbench — local projects and file tabs, frozen multi-file runs, experiment comparison as a persistent workspace, agent run control, and result provenance. It has grown from a “quantum demo page” into a frontend with project semantics.
  • Others: FlagGems-sglang #251 adds the remaining Qwen3.6 workload suite and the native SGLang adapter; Megatron-LM-FL #203 unifies multi-platform CI planning, affected unit-group routing, and evidence-based gating; FlagSparse #115 merges from the external repo NCIC-AlphaSparse/main; three OpenCourse commits establish course and version archives by release.

II. News Coverage and Ecosystem

2.1 Channel Verification: No New Hits in CSDN Column or HN Window; Zero Component-Level Items on the News Side (11-02)

Date: 2026-11-02　Sources: CSDN FlagOS Column · HN Algolia · Google News RSS

  • CSDN Column: After checking all 15 article links on the list page one by one, the two most recent pieces remain the 10-29 articles “The Savior of LLM Inference — One Plugin Handles Multi-Chip Adaptation” and “FlagOS Day0 Adaptation for MiniCPM-OCR & Jev: Covering 12 Platforms”; no new posts since 10-30. The page also carries replay entries for three conferences from 10-15 to 10-17.
  • HN Algolia: Eight query groups (FlagOS, FlagGems, FlagScale, FlagTree, BAAI, FlagQuantum, FlagSparse, flagos-ai) produced no direct title or body hits within a 5-day range; FlagOS and FlagGems each had 4 to 5 fuzzy hits, all unrelated to this topic.
  • Google News: 24 Chinese and English query terms returned 43 hits; after checking each headline, 0 were directly related to FlagOS component releases or chip adaptation. The hits were mainly Enflame and MetaX Q3 earnings and stock-price pieces, plus industry articles surfaced by broad queries such as “BAAI open source” and “domestic compute open-source software stack.” Component-name queries (FlagOS, FlagGems, FlagScale, FlagTree, FlagCX, KernelGen, etc.) continue to return zero hits, which is a long-standing norm.
  • BAAI Community: Both the homepage and search page are client-rendered; scraping can only confirm that no FlagOS entries appear on the homepage and cannot yield count-level conclusions. This channel continues to be recorded as “no new visible entries.”

2.2 Member Companies on the Capital Markets Side: Enflame and MetaX Q3 Reports Disclosed the Same Day (10-30)

Date: 2026-10-30　Sources: Enflame Q3 Report (CS.com.cn) · MetaX Q3 Report (The Paper)

  • Enflame (688801): Revenue for the first three quarters was RMB 2.332 billion, up 331.74% year on year; net profit attributable to shareholders was -RMB 754 million, with the loss narrowing year on year; Q3 single-quarter revenue was RMB 1.211 billion, up 395.34% year on year, with net profit attributable to shareholders of -RMB 122 million. R&D expenses were RMB 989 million, up 11.11% year on year, and Q3 R&D spending accounted for 28.83% of quarterly revenue. Contract liabilities were RMB 6.385 billion, up 1,745% from the beginning of the year. On the product side, it offers five cloud AI chips across four architecture generations, including AI accelerator cards, intelligent computing systems, and a programming software platform.
  • MetaX (688802): Revenue for the first three quarters was RMB 2.824 billion, up 128.45% year on year; net profit attributable to shareholders was RMB 639 million and non-GAAP net profit was RMB 259 million, both turning profitable year on year; Q3 single-quarter revenue was RMB 1.5 billion, up 367.12% year on year, with non-GAAP net profit of RMB 308 million, marking two consecutive quarters of single-quarter non-GAAP profitability. On the software side, it released MXMACA 3.9.0, and MXMACA build support code for MetaX GPUs has been merged into the open-source project LMCache; on the model side, it says it has completed Day 0 adaptation for models including GLM, Qwen, and Hunyuan.
  • Relationship to FlagOS: Both are FlagOS member companies, but this item concerns earnings and commercial progress, not FlagOS project developments. Of these, MetaX merging build support code into LMCache is the only action that lands directly in an open-source upstream; it aligns with FlagOS’s cross-chip goal but follows its own software-stack route. The corresponding work on the FlagOS side is still reflected in backend patches for FlagGems and Torch-FL (see 1.5 and 1.6).

2.3 Ecosystem Watch: World Internet Conference Wuzhen Summit Opens Today, Theme Focuses on Open Source and Openness (11-02)

Date: 2026-11-02　Sources: World Internet Conference · Wuzhen, Why “Open Source” · Xinhua: 2026 Wuzhen Summit to Focus on Open Source and Openness

  • Agenda: The 2026 World Internet Conference Wuzhen Summit runs from November 2 to 5 under the theme “Open Source and Openness, Co-Building and Sharing — Jointly Building a Community with a Shared Future in Cyberspace.” It features 27 sub-forums, and the main forum includes an “AI Wuzhen Dialogue,” with topics including building a global open-source AI ecosystem.
  • Relevance: This item is industry background and has no direct agenda link to the FlagOS project; it is included here because open-source AI system software is one of the summit’s main topics, and the “multi-chip unified software stack” in which FlagOS operates is the most concrete engineering form under that topic. Public coverage within the window did not involve any FlagOS-specific agenda, so it is not counted as a project development.

III. Deep Dive into Member Units and Backend Vendors

3.1 Hygon: DTK 26.10’s 21-PR Closeout from Base Image to Wheel (10-31 ~ 11-01)

Date: 2026-10-31 ~ 2026-11-01　Source: build-infra PR #1568 · build-infra PR #1577 · FlagCX PR #667 · FlagScale PR #1299

  • Image side: This round for Hygon was not a single-point bug fix but a completion of the entire DTK 26.10 delivery chain—the base image installs ROCm-aware Open MPI, the runtime image bootstraps first without pinning flagcx, then pins flagcx==0.14.1+dtk2610 back and releases that wheel separately. #1587 hands off the defects of the vendor Triton 3.7.0 to the vLLM line, effectively acknowledging that the problems with this Triton version will not be fixed on the FlagOS side.
  • Communication side: FlagCX #667’s unified packaging of Hygon bitcode and Device IR marks the first time Hygon has obtained a build artifact form on the communication library side consistent with other backends.
  • Training side: FlagScale’s under-review #1299 adds Hygon DCU hardware health sampling, while #1255 through #1257 add three types of diagnostics: missing CPU-side NCCL entries, collective communication and P2P mismatches, and internal hangs. #1299 together with the three diagnostics points to “whether problems can be localized when ten-thousand-card training fails.”
  • Attention side: FlagAttention #117 is under review, covering Hygon GLA backend optimization, with an update on 10-30.
  • Assessment: Hygon is the member unit with the most concentrated investment in this window—21 build-infra PRs plus communication and training-side patches, covering the three stages of “packaging—communication—training observability.” The cost of DTK 26.10 lies almost entirely in dependency resolution and packaging rather than the kernel itself, consistent with the conclusion from the previous issue’s first-round DTK 26.10 integration.

3.2 Ascend: Beyond the 910C Native Quantum Line, XCS Interface and Communication CI Advance in Parallel (10-30 ~ 11-01)

Date: 2026-10-30 ~ 2026-11-01　Source: FlagQuantum PR #958 · FlagGems PR #7257 · FlagCX PR #669 · FlagAttention PR #120

  • Quantum line: 910C native provider support and the ARM64 regression lane are the largest single item in this window (see 1.3). The accompanying FlagCX #686 is under review, handling Ascend device completion tracking and tensor keep-alive—a prerequisite for the quantum stack at the communication layer.
  • Operator line: FlagGems #7257 changes the Ascend topk custom operator invocation to tle_raw.call, aligning with the TLE dialect; this kind of interface convergence typically precedes large-scale operator migration.
  • CI line: FlagCX #669 is under review, building a CI pipeline for Ascend 910C and standardizing unit test orchestration; vllm-plugin-FL #663 is under review, switching the CI image to v0.28.0-ascend-ci and disabling graph-mode test cases on 910c.
  • Attention line: FlagAttention #120 is under review, adding Ascend GDN chunk forward and recurrent decoding.
  • Compiler line: flir #86 is under review, fixing the lowering path for Ascend structured permutation and compute-type FP32 dot.

3.3 Kunlunxin: Ten XPU3 Performance Patches Merged in a Single Day (10-30)

Date: 2026-10-30　Source: FlagGems PR #7988 · FlagGems PR #8024 · FlagGems PR #7955 · sglang-plugin-FL PR #155

  • Ten in one day: From #7954 to #8052, Kunlunxin merged ten FlagGems patches on 10-30 alone, covering elementwise, reduction, quantization parameter search, foreach fusion, and copy paths. Two common threads run through the implementations: first, favor direct pointers and single-issue to reduce intermediate allocations; second, give bf16 its own block width and kernel body rather than reusing fp16 parameters.
  • Precision-related: #8024’s fmod uses div_rn plus FMA residual for exact truncation, and #7955’s choose_qparams_optimized uses a single issue for neighborhood search—both examples of “tightening performance and numerical correctness simultaneously.”
  • Inference stack: sglang-plugin-FL #155 is under review, upgrading the Kunlunxin backend to SGLang 0.5.18; #251, merged on 10-31 on the FlagGems-sglang side, adds the Qwen3.6 workload suite.
  • Assessment: Kunlunxin is the only backend in this window appearing in the form of “batch performance patches,” with ten concentrated on the same day, indicating its team advances in batches rather than following the upstream cadence.

3.4 MetaX: FlagPrism Adds MACA Backend, Attention Side Adds Parallax and Hy3 (10-30 ~ 11-02)

Date: 2026-10-30 ~ 2026-11-02　Source: FlagPrism PR #18 · FlagAttention PR #119 · FlagAttention PR #122 · FlagGems PR #7589

  • Debugger: FlagPrism #18 was merged at 00:11 on 11-02, adding the MACA runtime adapter, build module, Python bindings, and tests. Previously MetaX was absent from FlagPrism’s backend list; with this PR, MetaX is on par with other backends in the multi-chip debugging chain.
  • Attention: #119 is under review, adding optimized Parallax prefill and decode for MetaX; #122 is under review, adding BF16 and INT8 Hy3 attention. Both are multi-chip implementations of new model structures rather than tuning of existing operators.
  • Operators: FlagGems #7589 was merged, fixing MetaX upsample_nearest2d_backward’s explicit scale mapping; Torch-FL #688 was merged, registering MetaX mm in FlagGems and relaxing the K rejected by Triton.
  • Inference stack: vllm-plugin-FL #635 is under review, adapting Qwen3.8-Flash-Next to MetaX; no new MetaX PRs were seen on the sglang side within the window.
  • Industry side: MetaX’s Q3 report disclosed the release of MXMACA 3.9.0 and the merge of build support code into LMCache; see 2.2.

3.5 Tsingmicro, Enflame, Iluvatar: TXDA into vLLM 0.28, GCU Graph Mode, FlagTree Version Pinning (10-30 ~ 11-01)

Date: 2026-10-30 ~ 2026-11-01　Source: vllm-plugin-FL PR #642 · vllm-plugin-FL PR #667 · Torch-FL PR #698 · FlagTree PR #1494

  • Tsingmicro: vllm-plugin-FL #642 is under review, adapting the eager TXDA backend to vLLM 0.28.0; sglang-plugin-FL #133 is under review, tolerating the qwen3_asr registration differences in transformers 5.15+ and fixing token_to_kv_pool leaks under multiple chunks; Torch-FL #704 was newly opened on 11-02, adding a registration set guard for Tsingmicro’s conf. The three PRs correspond respectively to “version tracking, long-run stability, and configuration correctness.”
  • Enflame: vllm-plugin-FL #665 is under review, migrating the CI image to flagos-dev and disabling graph-mode test cases on the v0.28 backend; #667 was newly opened on 11-01, adding native eager and graph execution for vLLM 0.32 for Enflame; FlagTree #1494 was merged, temporarily disabling the 03-topk test to stop the bleeding; FlagAttention #118 is under review, adding the parallax operator.
  • Iluvatar: Torch-FL #698 was merged, pinning Iluvatar CI’s FlagTree to 0.7.0—the version the chip actually runs; #697 was merged, routing baddbmm through the vendor GEMM; vllm-plugin-FL #664 is under review, adapting CI to v0.28.0 and updating the prebuilt image; sglang-plugin-FL #139 is under review, completing the empty-runtime adaptation for SGLang 0.5.18.
  • Commonality: All three vendors’ actions within the window center on “vLLM 0.28 migration,” with the difference being that Tsingmicro and Enflame are adding long-run and graph-mode support, while Iluvatar is pinning the version to the line its chip can actually run. Such work does not appear in operator performance tables, but it determines whether new versions can run stably on their own chips.

3.6 Moore Threads and Horizon Robotics: MUSA Side Adds Large-Model Attention, BPU Side Adds Operator Paths (10-30 ~ 11-01)

Date: 2026-10-30 ~ 2026-11-01　Source: FlagAttention PR #121 · sglang-plugin-FL PR #158 · Torch-FL PR #686 · Torch-FL PR #693

  • Moore Threads: FlagAttention #121 is under review, adding Moba attention and fixing an import defect; sglang-plugin-FL #158 is under review, supporting MiniMax-M3 on MUSA; FlagGems has a fix/musa-pinverse-frozen-reference-20261030 branch pushed but not merged. Moore Threads had no merged PRs within the window; all actions are on the under-review side.
  • Horizon Robotics: All three merged Torch-FL PRs center on BPU—conv, batch norm, and pooling go through the eager path onto BPU (#686), avg_pool is carried by an equivalent depthwise convolution (#693), and that form is then expressed via the functional API (#696). This is the only complete path in this window of “a new backend going from a single operator to a group of operators.”
  • Comparison: Moore Threads maintains the cadence of “adding attention per model,” while Horizon Robotics is filling in coverage of the basic operator family. The difference is that the former serves specific model launches, while the latter serves backend usability.

IV. Summary and Trend Observations

Date: 2026-10-30 ~ 2026-11-02　Sources: flagos-ai organization · FlagQuantum PR #958 · build-infra PR #1584 · Sina Tech: AI writes operators from a zero starting point

  • Pace: Within 72 hours, 17 repos were pushed, with 256 commits, 115 merges, and 275 under review. On a daily average basis, that is about 85 commits and about 38 merges, both below the 189 and 183 in the previous 24-hour window, showing a pronounced weekend effect. At the same time, the total under review rose 8.7%, with 146 newly opened, forming a combination of “lower output, higher backlog.”
  • Structure: FlagQuantum’s 148 commits and FlagQAI’s 42 together account for 190, or 74% of all commits; the mainline operators and compilation (FlagGems 16, Torch-FL 13, FlagTree 2) total 31, or 12%. The weekend output structure is sustained by the two new lines of quantum and agents, while the mainline depends on weekday vendor investment.
  • Release: The two image layers of 2.3.0 (Hygon DTK 26.10, AMD ROCm 7.2.4) have completed changelog, dependency pinning, and description refresh, but the release/2.3 manifest directory has still not been created, and the release-info site does not list 2.3.0. Following the 2.2 path, 2.3 still lacks two steps—manifest freeze and site launch—and the next work week is the key window.
  • Technical direction: In this window, the focus of the quantum stack shifted from “which constructs are supported” to “whether the resource accounts can be recomputed”—x * 256 dropping from 28 allocated qubits and 4 GiB to 19 and 8 MiB is a hard number that can be verified by regression. Together with the 167 hardware acceptance items on the 910C and the fail-closed regression lane, the quantum line is forming three checkable pieces of evidence—”semantics, resources, hardware”—which is closer to engineering usability than single-point performance.
  • Risk: FlagGems has 158 under review and 82 newly opened, the vast majority from KernelGen batch submissions, with 18 merges severely mismatched. If review throughput does not improve before the 2.3 freeze, either the release will be held up by the backlog, or a large batch of operators will bypass review into the mainline—both of which would weaken the “merged means verified” standard.
  • External reference (non-FlagOS project): On 10-30, ModelBest open-sourced the operator generation system ForgeKernel. In a public livestream, over 22.5 hours and 42 iterations on B200, it cut FlashMLA prefill from 1834 microseconds to 472 microseconds, 9.8% faster than DeepSeek’s 518 microseconds, and claims to top all three CANNBench leaderboards: A3, 950DT, and 950PR. It is in the same lane as FlagOS’s KernelGen, showing that “AI writing operators” has moved from paper metrics to public leaderboard competition and real framework integration; the lesson for FlagOS lies not in model capability but in whether the review and verification process can keep up with generation speed—exactly the problem exposed by this period’s FlagGems backlog.

Appendix

Source Verification Table

Source Verification Result
GitHub organization repositories (61 repos) 17 repos pushed within the window, all traced to commits or branch ownership
Commit search (committer-date) 256 commits, FlagQuantum leads with 148
Merged PR search 115 PRs, covering 14 repos
Open PR search 275 PRs, 146 newly opened, FlagGems accounts for 158
Release / Tag atomic verification (31 repos with Releases, 8 repos with Tags) No new Release objects within the window; FlagCX v0.14.1 has only a tag, the official Release remains v0.14.0
community release manifest release/2.3 not created (404), release/2.2 manifest still present
release-info publishing site No 2.3.0 image tag appeared
Google News RSS (24 Chinese and English queries, via proxy) 43 hits, 0 related to FlagOS components
HN Algolia (8 queries, 5-day range) No direct hits
CSDN FlagOS column Latest two posts dated 10-29, no new posts within the window
BAAI community hub.baai.ac.cn Homepage and search page are client-rendered, no new FlagOS entries found
Member company official websites (nine) Pages provide no parseable dates, no new posts after 10-30; financial reports based on CS.com.cn and The Paper coverage

Full Source List