Research window: 2026-11-18 06:00 ~ 2026-11-19 06:00 (Beijing time, 24 hours; continuing from the cutoff of the previous daily report).

Sources: On the GitHub side, all 65 flagos-ai repositories were checked, including 148 commits within the window, 149 merged PRs, 318 in-review updates within the window, and pushes to 19 repositories. In addition, Releases and Tags of 13 repositories, 24 groups of Google News queries, HN Algolia, and member organizations’ original sites were checked; see the appendix source list for details.


Index for This Issue

  • Today’s highlight: In a single day, three vendors were swapped out in the FlagOS 2.3 base layer — MetaX switched to maca3.10.1.11, Enflame’s two tops lines were delivered for the first time, and Moore Threads musa4.3.6 was fully retired across the chain (11-18)
  • Today’s highlight: BAAI and Qingmao Intelligence jointly released an end-to-end Token Factory solution; the commercial domestic Token Factory goes live on FlagOS (11-17~11-18)
  • I. Open-source project progress (GitHub activity)
    • 1.1 Organizational cadence: pushes to 19 repositories, 148 commits, 149 merges, densest between 15:00 and 16:00 (11-18)
    • 1.2 build-infra: three lines of the 2.3.0 base landed on the same day; all twenty merges were vendor swaps (11-18)
    • 1.3 FlagCX: v0.14.1 tagged; Enflame device attribute types now defined by the SDK (11-18)
    • 1.4 FlagAttention: four new attention operators in one day, two each from T-Head and Ascend (11-18)
    • 1.5 FlagGems: three additions for Kunlunxin; TLE pipeline fully switched to tl.range (11-18)
    • 1.6 FlagGems-vllm: fourteen SiliconFlow commits merged within one hour; Hygon W8A8 adds head_dim=256 (11-18)
    • 1.7 FlagTree: TLE tensor graph compatible with TMA; FlagGems-vllm 0.2.0 baseline on Hygon BW (11-18)
    • 1.8 FlagTrain: Kunlunxin and Enflame backends enter the training operator library on the same day (11-18)
    • 1.9 Quantum line: FlagQuantum wraps up QEC decoding with fourteen commits; FlagQAI builds a research workbench with twenty commits (11-18)
    • 1.10 Review queue: 318 updates, with FlagGems accounting for 64 (11-18)
  • II. News coverage and ecosystem
    • 2.1 Channel check: 24 query groups yielded 18 candidates; component names still have zero hits (11-18)
    • 2.2 Partner-side recap of the Zhongzhi FlagOS Conference: Mouxianfei claims to be the only designated simulation platform (11-18)
    • 2.3 CSDN column and BAAI community: no new technical articles within the window (11-18)
  • III. Member organization deep dive
    • 3.1 Moore Threads: Kimi K3.1 Pro training adaptation, 400-step convergence on dual-machine 16-GPU setup (11-17~11-18)
    • 3.2 MetaX and Enflame: the chip and SDK specifications behind the two vendor swaps (11-18)
  • IV. Summary

Today’s Focus: Three Vendors Swap Out the FlagOS 2.3 Base Layer in a Single Day — MetaX Moves to maca3.10.1.11, Enflame Ships Two tops Lines for the First Time, Moore Threads Retires All musa4.3.6 Chains (11-18)

Date: 2026-11-18 Source: #1894 · metax-maca3.10.1.11 · enflame-tops1.9.29

  • MetaX line swap: #1894 replaces the base-layer maca3.7.2.1 with maca3.10.1.11, targeting MetaX C550 chips, with driver 3.100.1 and base image ubuntu:24.04. The changelog states explicitly that this line supersedes maca3.7.2.1; the driver is installed in user space only, so the image builds without a card present, and the kernel driver stays on the host. #1895 adds the same-day changelog for the runtime layer, and #1896 adds the SDK-bundled ompi libraries to LD_LIBRARY_PATH.
  • Enflame first delivery: tops1.9.29 targets Zixiao C200 (S60) and tops4.1.6 targets L600; both lines are first-time deliveries. The delivery surface is identical: driver, runtime, CC, PTI, TOPSTX, TOPS ATen, EFML, ECCL, Triton GCU 1.0.20261116, and gculare-perftest, all based on ubuntu:24.04. /opt/triton_gcu is made a symlink to the vendor’s triton install location, which is where FlagTree’s toolkit.py looks for it.
  • Moore Threads retirement: #1902 retires the musa4.3.6 build layer from the base layer, and #1904 removes the same line from the application layer; #1905, covering the flagcx row, is still under review. The musa5.2.0 and musa5.3.0 lines are retained.
  • Version baseline: the stack version in configs.yaml is 2.3.0, with flaggems pinned at 5.4.0. The 2.3 schedule follows community #176: FEP merge deadline 11-13, feature freeze and rc0 on 11-30, rc1 on 12-07, rc2 on 12-14, final freeze 12-25, GA 12-30.
  • Observation: in a single day, two vendors changed generations and one retired a line, with every action landing in changelogs and Containerfiles rather than release announcements. The base-layer cadence is to get images building and passing acceptance first, and talk version numbers later; at this point there are still twelve days until rc0.

Today’s Focus: BAAI and Qingmao Intelligence Jointly Release End-to-End Token Factory Solution, Commercial Domestic Token Factory Goes Live on FlagOS (11-17~11-18)

Date: 2026-11-17 ~ 2026-11-18 Source: Leiphone · Huanqiu Tech · Investorscn

  • Event: On November 17, at the 2026 Qingmao AI Partner Conference and Token Factory Annual Meeting under the theme “Domestic Chips, Domestic Models, Independent Ecosystem,” the Beijing Academy of Artificial Intelligence (BAAI) and Qingmao Intelligence jointly released the “Domestic Token Factory End-to-End Inference Optimization Solution,” launching a commercial Token Factory built on Zhongzhi FlagOS. Leiphone published its report at 13:31 on 11-18, with Huanqiu Tech and Investorscn following the same day.
  • Ecosystem claims: Reports state that the FlagOS ecosystem now brings together 21 chip companies and 86 ecosystem partners, covering 35 AI chips. The nine operator libraries have accumulated more than 1,454 open-source Triton operators, of which nearly 800 achieve over 80% of native operator performance on at least 5 domestic chips. A single code repository covers backends for 20 vendors, and over the past six months it has repeatedly achieved same-day bring-up on 12 chips when new models are released.
  • The Ascend 910C numbers: For Zhipu GLM-5.3 and GLM-5.3-Flash, the Prefill side improves wave throughput by 3.41% at 81920 input and 16 concurrency, and reduces TTFT p50 by 4.94% at 256K input. On the Decode side, LI quantized write-back merges dynamic quantization, scale conversion, and HBM cache write-back into a single kernel, cutting the 910C microbenchmark from 18.03 microseconds to 2.04 microseconds. End-to-end ITL drops by more than 10%, and TPOT falls by over 8%.
  • Existing 910B fleet: BAAI Token Hub forms a Prefill/Decode disaggregated cluster from 48 910B cards — 32 cards split into 4 Prefill domains and 16 cards into 8 Decode domains — running GLM-5.3-Flash service with 1M context. It has run stably for two consecutive weeks, with TTFT P50 at about 2.3 seconds and TPOT P50 at 22 milliseconds. Cross-Prefill-domain Prefix Cache reuse raises the hit rate from 82.65% to 94.15%, and cuts median time-to-first-token during the reuse phase from 16.63 seconds to 3.10 seconds.
  • Cross-platform comparison: On T-Head 810E, FlagOS 2.2 improves DeepSeek-V4.1-Flash by 77% to 82%, supports 16-way concurrency in service with up to 80 concurrent requests, and achieves 136.2 tok/s decoding per request. On Hygon BW1100, GLM-5.3-Flash improves by more than 25%, with per-stream throughput exceeding 80 TPS under multi-concurrency.
  • Caveat on claims: The Huanqiu Tech version additionally states that on 910C, GLM-5.3-Flash overall output rises from twenty-plus TPS to fifty-plus TPS, logical service throughput is 2.04x pre-optimization, and unit compute cost is 19.4% to 27.6% lower than commercial API billing; it also claims the auto-tuner improves search efficiency by 298x. These figures come from vendor and media sources and have not been independently reproduced.

I. Open-Source Project Progress (GitHub Activity)

1.1 Organization Rhythm: 19 Repos Pushed, 148 Commits, 149 Merges, Peak at 15:00–16:00 (11-18)

Date: 2026-11-18 Source: flagos-ai organization · FlagGems · FlagGems-vllm

  • Totals: 19 of 65 repos had pushes within the window. Commit search returned 148 hits, 149 merged PRs, and 318 open PRs updated within the window.
  • Commit vs. merge distribution: The two are nearly one-to-one, indicating that day’s traffic largely went through squash merges. FlagGems 34, FlagGems-vllm 28, build-infra 20, FlagQAI 20, FlagQuantum 14, and the remaining eight repos totaled 32 (FlagTrain 9, FlagTree 5, FlagAttention 5, optest 4, FlagGems-sglang 4, FlagCX 3, Torch-FL 2, vllm-plugin-FL 1).
  • Time distribution: First commit at 07:13; the last was still progressing at 05:15 the next day. 16:00 peaked at 19 commits, with 15:00 and 10:00 at 18 each. After 20:00 it dropped to 11, and from 22:00 onward one or two per hour—wrapping up.
  • Two repos pushed but with no commits on main: community’s push came from the codex/fep23-flagtree-device-enablement-20261105 branch, corresponding to FEP-23’s FlagTree device enablement; docs had only IssueComment events in the window, with no source changes.
  • Observation: The merge rhythm formed a single peak at 15:00–16:00, coinciding with two batches of concentrated merges in build-infra and FlagGems-vllm. No new repos were created and no releases were published that day—only one tag on FlagCX.

1.2 build-infra: Three 2.3.0 Base Lines Land the Same Day, All Twenty Merges Are Line Swaps (11-18)

Date: 2026-11-18 Source: #1894 · #1898 · #1904

  • Composition: 20 merges. MetaX line 10 (6 line swaps and fixes, 4 changelog backfills), Enflame line 3, Moore Threads line 2, and the remaining 5 are one each for flagcx wheel bookkeeping, vLLM validation output annotation, and a vendor.sh fix to override PYTHONPATH, plus two 2.3.0 description refreshes.
  • MetaX sequence: #1894 swaps the base layer, #1895 backfills the runtime layer changelog, #1896 fixes the ompi library path, #1897 and #1899 backfill changelog dates, #1900 removes an outdated driver-version comment. Six in a row—the day’s most complete line swap.
  • Enflame sequence: #1886 pins FlagTree at 0.7.1 and removes the operator-fallback environment variable; #1887 relaxes the sglang warmup budget and drops three environment-variable pins; #1888 backfills runtime change dates for two lines.
  • Moore Threads sequence: #1902 retires musa4.3.6 from the base layer; #1904 removes that line from the application layer. The rationale for the application-layer commit is in the message: no image needs that line anymore, and keeping it only adds a blank cell to the config matrix.
  • Bookkeeping and release: #1883, #1889, #1901, and #1903 write 2.3.0 into each layer’s image description and refresh the descriptions; #1898 and #1876 record the flagcx wheel release.
  • Observation: None of the twenty is a new feature. The base layer’s work that day was aligning “which line is in, which line is out” across configs and docs, so acceptance doesn’t read stale wording.

1.3 FlagCX: v0.14.1 Tagged, Enflame Device Attribute Types Now Defined by the SDK (11-18)

Date: 2026-11-18 Source: v0.14.1 · #721 · #723

  • Tagging: v0.14.1 was tagged at 11:05 Beijing time, pointing to commit 268d6aa2. The tag was produced by #724, the same fix as #723 on the 0.14.0-rc2 branch.
  • Fix content: The Enflame platform no longer guesses device attribute types by name, instead using the types provided directly by the SDK. A small change, but it determines whether new vendor fields can be handled at compile time.
  • Platform capability: #721 adds IXCUDA streams, host memory registration, and DMA-BUF support for Iluvatar—a functional fill-in for the communication library on that platform.
  • Rhythm: v0.14.0 was the formal release on 09-28, and v0.15.0.dev0 opened as early as 10-18. v0.14.1 is backported from the 0.14 rc2 branch—a maintenance-line patch, not a feature release.

1.4 FlagAttention: Four New Attention Operators in a Day, Two Each for T-Head and Ascend (11-18)

Date: 2026-11-18 Source: #143 · #134 · #133

  • Two for T-Head: #143 supports Inkling FA4 relative-position attention; #134 adds a Log-Linear Attention operator. Both landed around 14:19 and 14:26, aimed at bringing attention variants from new papers into the operator library.
  • Two for Ascend: #133 adds a fused attention residual forward; #135 adds chunked GLA forward and backward backends. GLA is a linear-attention variant, and the chunked implementation saves memory on long sequences.
  • One doc commit: #150 replaces an outdated TLE pipeline reference with tl.range, the same migration as the same-named one in FlagGems that day.
  • Observation: Merges in this library over the past month-plus were mostly small fixes; four operator variants in one day is the most concentrated feature expansion in recent memory. What the four share: “new model structures land first, performance tuning later.”

1.5 FlagGems: Three Kunlunxin Fill-Ins, TLE Pipeline Wholesale Switch to tl.range (11-18)

Date: 2026-11-18 Source: #8896 · #8894 · #8456

  • Three for Kunlunxin: #8896 adds coverage for linalg_inv, inv_ex, lu_solve, and polar; #8895 adds backends for index_reduce and histogramdd_from_bin_cts; #8894 adds real XPU kernels for the inverse, isreal, and kaiser_window series. All three merged between 18:25 and 18:27.
  • Triton upgrade regression: #8846 fixes Kunlunxin operator regressions after the Triton upgrade, forming the same fix line as the three above.
  • TLE migration: #9373 replaces the removed TLE pipeline iterator on the T-Head PPU side with tl.range; #9392 migrates cross attention off the TLE pipeline as well. Together they show the compiler-side TLE pipeline abstraction has been withdrawn, and upper-layer operators must follow.
  • Operators and backends: #8456 fixes attention padding V to the block boundary on KV tail blocks; #9342 adds conv2d to the T-Head backend; #9372 adds a dedicated SDPA forward for Tianshu; #9066 and #9121 migrate in Enflame’s test suite.
  • Six CI commits: #9358, #9400, #9418, #9437, #9448, and #9461 focus on compiler pinning for backend tests, shard counts, and triton shadow-directory cleanup—one pipeline cleanup effort.
  • Observation: Of the day’s 34 merges, Kunlunxin and T-Head each take an end, with a compiler-generation migration in between. The operator library’s daily work is shifting from “adding operators” to “keeping up with compiler interface changes.”

1.6 FlagGems-vllm: Fourteen SiliconFlow Merges Within an Hour, Hygon W8A8 Adds head_dim=256 (11-18)

Date: 2026-11-18 Source: #1283 · #1173 · #1267

  • Fourteen from SiliconFlow: 14 of the 28 merges came from SiliconFlow, concentrated within an hour and a half from 15:06 to 16:41. The three backend-rollout PRs are #1283 triton_ihc_pre across six backends, #1223 fused_kimi_k3_mla_kv_concat across five backends, and #1224 hc_down_silu across six backends, plus #1262 gumbel_sample. The other ten target cold-compile and CPU-reference contention: #1260, #1263, #1272, #1273 for Ascend cold compile; #1251 for Ascend and T-Head fused gate timeouts; #1264, #1269, #1270 to ease CPU reference thread contention; #1271 to reuse grouped Gemma RMSNorm across row counts; #1220 to fix the missing baseline in pad_indexer_heads.
  • Hygon: #1173 fixes the probability rounding issue in W8A8 FA2 and supports head_dim=256, raising the attention head-dimension ceiling by one notch.
  • T-Head: #1267 optimizes mhc_shifted_post_pre on PPU.
  • Two for MetaX: #1279 adds a MetaX backend for vLLM’s ple_conv; #1289 turns candidate block selection into a self-contained pure-Triton block top-k.
  • Others: #1151 adds an upstream-compatible ple_conv (for qwen4); #1291 makes grouped_topk support per-token padding masks; #1232 adds Xiwang’s stride-aware Flash KV cache write; #1051 adds an NVIDIA-side TLE bad-word mask operator; #1116 adds compute_topk_scores; #1266 fixes Ascend pack_seq and unpack_seq; #1301 bumps the vendor-side vLLM test dependency to 0.32.0; #1293 adds a vllm_api field to the 285B model shape benchmark; #1249 fixes NVIDIA’s KDA, bincount, and Marlin checks, later rolled back by #1303.
  • Observation: The day’s 28 merges center on “one operator across multiple backends” and “cold-compile time”—both pre-launch engineering issues. One contributor supplying half, concentrated within an hour, indicates a batch submission formed in one go rather than scattered fixes.

1.7 FlagTree: TLE Tensor Graph and TMA Compatibility, FlagGems-vllm 0.2.0 Baseline on Hygon BW (11-18)

Date: 2026-11-18 Source: #1539 · #1538 · #1673

  • Tensor graph: #1539 exposes tensor graph publication and descriptor views on the NVIDIA side—a step in TLE’s extension toward low-level description capabilities; #1538 preserves the async copy contract and TMA call compatibility. Both merged after 16:00, two sides of the same TLE rework round.
  • Three CI commits: #1673 enables the FlagGems-vllm v0.2.0 baseline on Hygon BW (hcu3.6); #1685 disables a failing TLE-Raw test case in the hcu workflow; #1678 bumps git-auto-commit-action from v4 to v7.
  • Observation: Running the v0.2.0 baseline on Hygon BW means this accelerator card enters the same tier of regression coverage as Ascend and T-Head. Disabling a failing test case is a temporary measure, indicating that path is still converging.

1.8 FlagTrain: Kunlunxin and Enflame Backends Enter the Training Operator Library the Same Day (11-18)

Date: 2026-11-18 Source: #99 · #91 · #88

  • Backend expansion: #99 adds both Kunlunxin and Enflame backends at once—the key commit expanding this library from two backends to four.
  • Operators: #91 updates blocked_flash; #85 updates fused scaled masked softmax forward and backward; #88 adds a portable fused_bias_gelu_impl forward; #79 updates fused gated norm and fixes a baseline comparison error; #83 aligns quick_gelu with Megatron on extreme negative inputs.
  • Engineering: #97 adds config() to the baseline and fixes backward profiling.
  • Observation: All nine FlagTrain commits center on training-side fused operators, contrasting with the inference-side FlagGems-vllm. Both libraries added backends on the same day, showing multi-chip coverage has extended from inference to training.

1.9 Quantum Line: FlagQuantum’s Fourteen Close Out QEC Decoding, FlagQAI’s Twenty Build a Research Workbench (11-18)

Date: 2026-11-18 Source: #1213 · #1212 · #116

  • FlagQuantum error-correction side: #1213 adds native batched decoding with weights; #1207 adds native packed batched decoding; #1212 pushes the color-code exact-distance proof to d=15; #1208 validates the full Surface and Repetition distance tables.
  • FlagQuantum compiler side: #1214 moves pipeline stages into their respective packages; #1211 adds standard paths for offline routing artifacts; #1204, #1209, #1210, and #1221 add validation for various artifact projections; #1218 records approval of auto-routing defaults.
  • FlagQAI: 20 merges, numbered consecutively from #97 to #116, all on the research workbench and Agent—grouping research history by project, comparing candidates against adopted baselines, packaging research results into downloadable artifacts, and enforcing backend guards on task submission.
  • Positioning check: FlagQAI’s README explicitly states its current boundary: what’s implemented is local circuit simulation and a two-qubit variational path with one fixed ansatz; high-level “compiling intent into circuits” is a product direction, not an existing capability.
  • Observation: The quantum line had 34 commits that day, nearly a quarter of the whole organization. FlagQuantum is engineering error correction, FlagQAI is building an interface for the research workflow—the division of labor is clear.

1.10 Open Queue: 318 Updates, FlagGems Accounts for 64 (11-18)

Date: 2026-11-18 Source: flagos-ai · #9490 · #850

  • Distribution: FlagGems 64, FlagGems-vllm 14, FlagTree 4, FlagQAI 4, FlagQuantum 3, vllm-plugin-FL 3, FlagCX 2, Torch-FL 1, sglang-plugin-FL 1, FlagGems-sglang 1, build-infra 1, docs 1, FlagTrain 1.
  • Moore Threads batch submission: Six variants of _foreach_clamp_min and _foreach_clamp_max (#9475 to #9480) were submitted at the same moment—a batch operator rollout.
  • Framework side: Torch-FL #850 submits Enflame native host and full-model validation; FlagGems #9488 submits quant_batch_matmul_mxfp8; #9483 and #9481 fix Kunlunxin’s sparse index metadata and upsampling backward.
  • Observation: The open-review volume is more than twice the merge volume, and the queue hasn’t shortened. FlagGems alone accounts for a fifth, indicating the operator library’s submission rate still outpaces review throughput.

II. News Coverage and Ecosystem

2.1 Channel Verification: 24 Query Sets, 18 Candidates, Still Zero Hits on Component Names (11-18)

Date: 2026-11-18 Source: Google News RSS · HN Algolia

  • Query combinations: 24 Chinese and English query terms, covering FlagOS and seven component names, BAAI and member organizations, and industry terms. 18 candidates were hit within the window.
  • Zero-hit portion: Component-name queries for FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, KernelGen, and Open3D-PIMC all returned zero hits, and FlagOS and BAAI FlagOS on the English side were likewise zero. This state has persisted across multiple windows.
  • HN: The three queries FlagOS, FlagGems, and FlagTree returned zero hits on HN Algolia.
  • Filtering: Of the 18 candidates, about ten were ETF quotes, individual stock research reports, reposted commentary, and gambling-related SEO, all removed in bulk by title. The remaining five entered the main text or appendix.
  • Observation: Public coverage of FlagOS centers on industry narratives such as “ecosystem collaboration” and “Token factory,” while component-level version and performance news still flows only through GitHub and community channels.

2.2 Partner-Side Recap of the Zhongzhi FlagOS Conference: Mouxianfei Claims to Be the Sole Designated Simulation Platform (11-18)

Date: 2026-11-18 Source: CNFOL · Earlier conference agenda

  • Timing check: This article, dated 11-18 at 18:13, covers the 2026 AI Open Computing Conference and Zhongzhi FlagOS Technology Conference held October 17-18 at the Exhibition Center of the Zhongguancun National Independent Innovation Demonstration Zone in Beijing, under the theme “One Stack Throughout, Boundless Intelligent Computing,” with 1 main session, 6 sub-forums, and hands-on workshops. It is a partner-side recap published one month after the conference ended, not a new event.
  • Mouxianfei’s account: Beijing Mouxianfei Technology Co., Ltd. states that it is the sole partner in the physical simulation domain and the sole designated simulation platform partner enterprise of FlagOS under the Zhongzhi FlagOS acceleration program project guided by the Beijing Municipal Science and Technology Commission and the Zhongguancun Administrative Committee. Co-founder and chief architect Sun Jiaqi presented at the embodied intelligence session.
  • Technical details: The MotrixSim physics engine, MotrixLab training platform, and MotrixGen data engine form a fully self-developed stack, and the UniLab open-source academic collaboration project has iterated to 2.0. The article states that in specific post-training scenarios, using roughly 2 to 3 times the volume of real-machine data in simulation data can match the success rate of post-training on pure real-machine data, and that the simulation training framework runs more than 3 to 5 times faster than other open-source simulators.
  • Ecosystem-side information: The same article mentions that FlagOS is carrying out China-Africa entrepreneurship cooperation with the United Nations Development Programme, conducting global AI computing teacher training covering more than 30 countries and 50 teachers, and that the United Nations has established an open computing working group with FlagOS as its technical foundation.
  • Observation: “Sole designated” is the vendor’s own account; no corresponding announcement appears on official FlagOS channels, so it is recorded here as the partner’s stated position.

2.3 CSDN Column and BAAI Community: No New Technical Posts Within the Window (11-18)

Date: 2026-11-18 Source: qq_41610763 · BAAI Community

  • CSDN column: The most recent visible post in the FlagOS intelligent computing system software stack column is the August 21 call for submissions for the SGLang operator optimization challenge; there were no new posts within the window.
  • BAAI Community: Several pushes within the window were model- and leaderboard-related content (Tencent Hunyuan Flash, materials science agent MatOS, Wujie PhysCapBench, Gong Ke’s speech), with no direct relation to FlagOS, and were not included.
  • Observation: On the day, GitHub was the only remaining technical channel, which also explains why the daily report focuses on commits and change records.

III. Deep Dive on Member Units

3.1 Moore Threads: Kimi K3.1 Pro Training Adaptation, Dual-Machine 16-Card Convergence in 400 Steps (11-17~11-18)

Date: 2026-11-17 ~ 2026-11-18 Source: Yicai · China Business Journal reprint · EEFocus technical article

  • Event: After Moonshot AI open-sourced the 2.8-trillion-parameter Kimi K3.1 Pro, Moore Threads completed training adaptation based on MTT S5000 intelligent compute cards and the MUSA software stack. It built a native FSDP2 multimodal training path, connecting the full chain from model loading and shard scheduling to gradient synchronization. Yicai and National Business Daily first reported on the evening of 11-17, and China Business Journal added technical details on 11-18.
  • Validation configuration: The trimmed training setup used a 13-layer language backbone, 16 experts (top-8 routing), and the full 27-layer vision tower, with 16 MTT S5000 cards across two machines, starting training from random initialization on MedPix-VQA image-text data. Sixteen processes completed 400 consecutive optimizer updates; training loss converged from 8.9365 to 2.1802, validation loss dropped from 6.5297 to 4.6873, and peak memory was 44.80 GiB.
  • Engineering highlights: Completed forward and gradient implementations for newly added operators, and integrated expert dispatch and combine into a unified computation graph; for parallelism, used a hybrid of FSDP2 and EP with parameter prefetching and gradient communication scheduling; on the loss side, used Chunk Loss blocking (block length 256) to control memory for the large vocabulary.
  • Scope caveat: China Business Journal explicitly noted that training adaptation does not equal pretraining the model on that chip; the two must be described separately. Moore Threads previously completed from-scratch training of the 021 scientific foundation model (236 billion parameters) and, together with BAAI, completed full-process training of the embodied brain model RoboBrain 2.5.
  • Observation: The technical substance of this news lies in “training adaptation” rather than “inference usability,” and the validation scale is a trimmed configuration. It shows that domestic cards are advancing their engineering work on trillion-level models, but there is still a gap to full-scale pretraining.

3.2 MetaX and Enflame: The Chip and SDK Scope Behind Two Line Switches (11-18)

Date: 2026-11-18 Source: base layer changelog directory · metax-maca3.10.1.11 · enflame-tops4.1.6

  • Chip correspondence: The new MetaX line targets C550, while the two Enflame lines target Zixiao C200 (S60) and L600 respectively. The two previously retired Enflame lines had already been replaced with transaction lines on 11-17; that day marked the first delivery of the newly built lines, closing out the same generation switch.
  • Driver strategy: MetaX makes the driver user-space-only installation, with no kernel modules installed inside the container. This aligns with NVIDIA’s approach; the benefit is that images can be built without a card, at the cost that the host driver version must match the image.
  • SDK component surface: The component lists for the two Enflame lines are completely identical, indicating that the difference between the two lines lies only in chip and stack versions, not in capability scope. Triton GCU is 1.0.20261116 for both, i.e., the build from November 16.
  • Directory consolidation: The maca3.7.2.1 image directories in the base layer and runtime layer have been deleted, leaving only historical changelogs; the directories of the old Enflame lines are likewise no longer in the tree. This “delete directories, keep records” approach keeps the build matrix clean while preserving traceability.
  • Observation: The base layer’s generation-switch scope is “one line per card.” Three vendors completed this on the same day, indicating that the platform list for the 2.3 cycle is basically finalized, and what follows is acceptance before rc0.

IV. Summary

  • The base layer was the day’s main thread: FlagOS 2.3’s base and runtime layers completed the MetaX line switch, the first delivery of both Enflame lines, and the retirement of the old Moore Threads line within one day; none of the twenty merges introduced new functionality. With twelve days left until rc0 on 11-30, the platform list is converging.
  • The operator layer is doing two things: FlagAttention added four attention variants at once, while FlagGems-vllm rolled out fused operators backend by backend. The former expands capability boundaries; the latter fills in coverage density.
  • Compiler interfaces are changing generations: The TLE pipeline iterator was removed, and FlagGems, FlagGems-vllm, and FlagAttention all switched to tl.range on the same day. This is a cross-repository interface alignment.
  • The quantum line’s share rose: FlagQuantum and FlagQAI totaled 34 merges, nearly a quarter of the day’s merge volume. The error-correction side is being engineered, while the research side is being interface-ized.
  • Only one main thread in public reporting: The Token factory solution from BAAI and Qingmao Intelligence is the only industry news with substantive content; the rest are partner-side retrospectives and reprints of old news. Component-level version news still only goes through GitHub.
  • To watch: One flagcx line for the old Moore Threads line is still under review; one test case is disabled when the FlagGems-vllm 0.2.0 baseline is enabled on Hygon BW; 2.3’s GA is set for 12-30, with three intermediate milestones: rc0, rc1, and rc2.

Appendix: Source Verification Table and Complete Source List

Source Verification Table

Source Verification Result
GitHub org: flagos-ai Full verification of 65 repos; 19 repos pushed within the window; 148 commits, 149 merged PRs, 318 in-review updates
GitHub Release / Tag Atomic verification of 13 repos; only one new tag, FlagCX v0.14.1; no new Releases
Google News RSS 24 Chinese and English query sets; 18 candidates; about ten excluded; five included in the main text
HN Algolia Three queries for FlagOS, FlagGems, and FlagTree returned zero hits
CSDN FlagOS column Most recent visible post dated August 21; no new posts within the window
BAAI Community Posts within the window were model- and leaderboard-related, not directly related to FlagOS
Member company original sites Moore Threads Kimi K3.1 Pro training adaptation (Yicai, NBD, China Business Journal)
community 2.3 roadmap community #176, target GA 12-30, no updates that day
Quantum line documentation FlagQAI README clearly defines current capability boundaries; included as per the original text

Complete Source List