FlagOS Daily Intelligence Report (2026-08-20)
Research window: 2026-08-19 10:19 ~ 2026-08-20 10:18 Beijing Time Sources: GitHub (org: flagos-ai), Google News aggregation (Chinese and English), HN, vendor official sites, etc. (see appendix)
Index
- I. Open-Source Project Progress (GitHub Activity)
- 1.1 SUNRISE of Xiwang Xinke completes full inference pipeline; build-infra follows up intensively with 12 commits (08-19~08-20)
- 1.2 vllm-plugin-FL adds Kunlunxin backend; Qwen3.6-35B-A3B/27B validation passed (08-19)
- 1.3 Torch-FL completes three-chip functionality on the same day: T-Head PPU AMP done, Moore Threads MUPTI, Ascend AMP extension (08-19)
- 1.4 FlagScale adds Ascend and MetaX build/test CI (08-19)
- 1.5 FlagBLAS L2 operator systemization completed; AriesMTR series PRs merged intensively (08-19)
- 1.6 sglang-plugin-FL Ascend NPU end-to-end validation (08-19)
- 1.7 Other repository activity: FlagTree / FlagGems / FlagGems-Experimental / TransformerEngine-FL / FlagCX / docs (08-19~08-20)
- II. News Coverage and Ecosystem
- 2.1 Overall assessment of news coverage
- 2.2 Ecosystem reference items: Bosch Journey 6B ten-million-unit designation, two consecutive Hygon ecosystem partnerships, Moore Threads Hong Kong stock, MetaX space-based intelligent computing, Cambricon lock-up expiry proceeds received (08-19)
- 2.3 Trends over the past 3 days
- III. Deep Dive into Member Organizations
- IV. Summary
- Appendix: Complete Source List
1. Open-Source Project Progress (GitHub Activity)
Window Overview: Of the 53 repos in the org, 14 had pushes during the window; commit search returned 65 commits within the window (build-infra 24, FlagBLAS 19, leading the pack), with no official GitHub Release. Headline of this window: build-infra completes the full SUNRISE inference pipeline for Xiwang Xinke—the FlagTree sunrise wheel production line was formalized and passed end-to-end validation for both vLLM versions (0.24.0/0.20.2), along with a fix for an online decode hang; additionally, vllm-plugin-FL adds Kunlunxin backend support for the first time, Torch-FL completes three-chip functionality on the same day, and FlagScale adds Ascend/MetaX CI.
1.1 build-infra: Full SUNRISE Inference Pipeline for Xiwang Xinke (Window Headline, 12 Commits)
Source: build-infra commits
- Wheel production line formalized (8/19 15:26 #442): Added the
packaging/flagtree/sunrisebuild mechanism—ubuntu 22.04 old-glibc base (same pattern as the metax builder), pre-installed sunrise dependencies (LLVM 22 / sunriseTritonPlugin_v0.6.0.4 / triton toolkits), building the FlagTree wheel from source withFLAGTREE_BACKEND=sunriseand applying version gating; a corresponding CI target was also added. On the same day, #441 completed end-to-end validation of vllm-sunrise 0.24.0 on TANGRT 1.2.0 (including DOCKER_RUN_FLAGS for the ptpu build). - Build troubleshooting series (8/19 17:16-20:25 #445-#450): clang/lld replacing gcc for the build, ld.lld toolchain check, building cp310 wheel with system python3.10, switching pip to the Alibaba Cloud mirror (runner cannot reach pypi.org directly), restoring $ORIGIN RUNPATH with the tang alias attached.
- Online issue fixes and dual-version re-verification (8/20 07:39-09:01 #451/#459/#461): Rebuilt the sunrise FlagTree wheel to fix the decode hang (recorded in vllm 0.20.2 §2.9 and 0.24.0 §11.5); subsequently completed sunrise 0.24.0 vllm-repack end-to-end (#459) and 0.20.2(F) re-check (#461).
- Other commits in the same repo: metax flagtree training E2E and merged-wheel usage change notes (#435), megatron-rl app environment pinned with TORCHINDUCTOR_COMPILE_THREADS=1 (#443), NVIDIA flash-attn source build workflow (#454), NVIDIA cuda13.3 four-scenario dual-compiler validation notes (#455), app image prefix unified to flagos-app with standardized tag naming (#457/#462/#463).
Interpretation: After Xiwang Xinke SUNRISE completed its FlagTree backend dependency update on 8/18 (#999), this window moved directly into a deep delivery phase of “wheel production line + inference E2E + online issue fixes“—from build base selection to mirror source adaptation to decode hang fixes, the end-to-end engineering completeness is the most thorough example among recent domestic backend integrations (previously only T-Head PPU reached similar depth). Meanwhile, the NVIDIA-side megatron app image mechanism (flash-attn source build, cuda13.3 dual-compiler matrix, tag naming conventions) is also being refined in parallel, with the training/inference image system trending toward unification.
1.2 vllm-plugin-FL: Adds Kunlunxin Backend (#268)
Source: PR #268 (merged 8/19 15:16)
Based on vLLM 0.20.2 (empty-library adaptation), adds Kunlunxin backend support to vllm-plugin-FL, validated with Qwen3.6-35B-A3B and Qwen3.6-27B, with tests covering high-concurrency text and image inputs (TP=4, max-model-len 16384, integrating FLAGCX_PATH and USE_RESHAPE_AND_CACHE_FLASH). Kunlunxin had just merged its first batch of operators into the KernelGen production line on 8/18 (digamma, arctan minimax, bernoulli), and entered the inference plugin line the very next day—the domestic backend matrix on the inference side gains another seat (previously covering Enflame, Iluvatar, Moore Threads, Ascend, etc.).
1.3 Torch-FL: Three-Chip Functionality Completed on the Same Day (6 Commits)
Source: Torch-FL commits
- #140 (8/19 13:38) T-Head Zhenwu PPU AMP integration complete—the automatic mixed-precision pipeline for the PPU backend is now complete.
- #136 (8/19 13:13) Moore Threads MUSA MUPTI profiler integration—the second chip vendor’s profiler integration, following Ascend MSPTI profiler on 8/18.
- #135 (8/19 14:05) Ascend AMP and dtype support extended.
- #132 (8/19 13:41) Fixed torch.compile fullgraph compilation on flagos; #133 (14:45) FlagGems index_select full-route box; #138 (12:51) Added FlagGems integration guide documentation.
Interpretation: Continuing the 8/18 “multi-chip functionality completion” theme, this window saw three chips (T-Head/Moore Threads/Ascend) landing AMP and profiler capabilities on the same day, further deepening coverage at the training plugin layer.
1.4 FlagScale: Adds Ascend and MetaX Build/Test CI (3 Commits)
Source: FlagScale commits
- #1250 (8/19 11:15) Added Ascend and Metax build and test CI workflows—the training framework CI matrix formally extends to two domestic chip vendors.
- #1263 (8/19 13:00) Configure FL backends before TP overlap initialization; #1262 fixed chat template return type compatibility.
1.5 FlagBLAS: Systematic Level-2 Operator Completion (19 Commits, AriesMTR PR Series)
Source: FlagBLAS commits
- Level-2 operator family batch advancement (AriesMTR consecutive PRs #71-#81): SYMV/HEMV (8/19 11:20), SPMV/HPMV (12:59), GER/SYR/HER (13:55), SPR/HPR (14:54), SYR2/HER2 (15:14), GBMV/HBMV/SBMV (16:00) support merged successively, with SPR2/HPR2 benchmarks updated in sync (10:39).
- Engineering governance: L2 benchmark naming fixes (17:26), fault tolerance to continue running on pytest collection errors (#81, 8/19 18:05).
Interpretation: FlagBLAS’s Level-2 sparse/banded matrix operation family was systematically completed within a single day, continuing the L2 benchmark governance cadence of 8/18 (#68/#70), with the library’s operator coverage rapidly converging toward the BLAS specification.
1.6 sglang-plugin-FL: Ascend NPU End-to-End Validation (#57)
Source: PR #57 (merged 8/19 11:13)
[e2e] Feat/ascend npu e2e—the Ascend NPU end-to-end pipeline for the sglang plugin line is merged. Following the Iluvatar backend (#34) and Moore Threads MUSA e2e (#54) on 8/18, Ascend is now complete on the sglang side.
1.7 Other Repository Activity
- FlagTree (2 commits): #946 (8/19 10:32) treats TLE extract_tile/insert_tile as layout anchors; #1014 (14:55) fixes benchmark tests and tle build options.
- FlagGems (2 commits): #5592 (8/19 11:56) i0 path avoids unnecessary x.to() temporary tensors; #5628 (17:30) updates FlagTree and vendor docker images.
- FlagGems-Experimental (2 commits): #286 (8/19 14:13) [KernelGen][Iluvatar] adds special_modified_bessel_k0 and index_copy_ dedicated operators (Iluvatar); #315 (12:35) exports split_with_sizes.
- TransformerEngine-FL (1 commit): #95 (8/20 10:02) exposes backend ops needed for TP communication overlap.
- FlagCX (1 commit): #548 (8/20 09:36) updates version metadata to reflect the latest release (documentation).
- docs (4 commits): automatic model list sync (modelscope / huggingface / aihuanxin, 8/19 11:00-11:01).
- Quiet repos: FlagPerf, FlagAttention, FlagQuantum, FlagOS-Robo, KernelGen, FlagTensor, Megatron-LM-FL (main branch), FlagSparse, etc. had no new commits on the main branch during the window.
II. News Coverage and Ecosystem
2.1 Overall News Assessment
Google News returned zero direct hits within the window for FlagOS/FlagGems/FlagScale/FlagTree/FlagPerf/FlagCX/KernelGen component keywords (consistent with 8/19); HN had only 1 irrelevant hit (a “TokenBlast” Claude Code tool post, excluded). The “BAAI open source” query remains dominated by gambling/SEO noise (excluded in full). Component-level news remains quiet; substantive ecosystem items center on member organizations’ commercial developments (see 2.2 and Chapter III for details).
2.2 Ecosystem Reference Items (within the 8/19 window)
| Date | Event | Source |
|---|---|---|
| 8/19 | Bosch’s fourth-generation multi-function camera MPC4, based on Horizon Journey 6B, wins a ten-million-unit-class production nomination: covering 40-plus vehicle models, with the first nominated project entering mass production in 2026 Q3; Journey 6B has secured nominations with over 25 automakers, covering more than 50 models, with lifetime nomination volume exceeding 20 million units (Horizon is a FlagOS member, BPU architecture) | Guandian, Jiemian/East Money, Economic Observer |
| 8/19 | Hygon Information ecosystem partnerships x2: signed a strategic cooperation agreement with Tianji Data; established a joint innovation laboratory with Koal Software | Sina Finance · Tianji Data, Sina Finance · Koal Software |
| 8/19 | Moore Threads “heads to Hong Kong for another major test”: second push for a Hong Kong listing after nearing break-even (follow-up to the 8/18 “planning Hong Kong IPO”) | QQ News |
| 8/19 | In-depth report on MetaX Tianji intelligent computing: retrospective on the “Yunjian MetaX” intelligent computing satellite (Jitianxing A-04) — launched into orbit on 7/24 by Lijian-1 Y15, carrying an intelligent computing payload based on MetaX GPUs | Sohu |
| 8/19 | First tranche of Cambricon equity incentives for 124 employees vested and paid out: total of approximately RMB 690 million, averaging about RMB 5.57 million per person (formal implementation of the vesting reported on 8/18) | Sohu |
Excluded items: Horizon’s “Journey 7 finalizes target compute, cockpit-driving integrated product named Xingkong” is old news from 2026-03 republished via gnews (Autohome repost at 8/19 20:32, excluded); Unitree’s STAR Market listing (not a member organization); blanket chip stock selloff and ETF fund share reports; news about the game series “Horizon”; the general AI article “BAAI Community · Alibaba Agent Context” (no substantive connection to FlagOS).
2.3 Recent 3-Day Trends (corroborated by GitHub org pushed)
| Date | Active repos | Main content |
|---|---|---|
| 8/18 | FlagGems-Experimental(80), build-infra(18), FlagGems(14), FlagBLAS(12), FlagTree(6), Torch-FL(5), sglang/vllm-plugin-FL and 15 repos total | KernelGen batch shipping across all vendors (9 backends merged the same day); FlagTree adds SpaceMiT backend (14 vendors); sglang/vllm plugins go multi-chip |
| 8/19 | build-infra(24), FlagBLAS(19), Torch-FL(6), FlagScale(3), docs(4), vllm-plugin-FL, FlagGems, FlagTree, FlagGems-Experimental, sglang-plugin-FL and 14 repos total | SUNRISE wheel production line launched + vllm 0.24.0 E2E; Kunlunxin joins the vllm plugin line; FlagBLAS L2 operator family completed; Torch-FL three-chip AMP/profiler; FlagScale Ascend/MetaX CI |
| 8/20 (tail of this window) | build-infra(8 commits), TransformerEngine-FL, FlagCX | SUNRISE decode hang fixed + dual-version re-verification of 0.24.0/0.20.2 completed; megatron app image mechanism refined (flash-attn source build, cuda13.3 four-scenario validation, tag conventions); TP comm overlap ops exposed |
Trend assessment: Following 8/18’s “KernelGen parallel shipping across all vendors,” the focus of this window shifted to deep backend implementation and inference pipeline integration — Xiwang Xinke’s SUNRISE, with a complete model of “wheel production line + E2E + online issue fixes,” has become the fastest-adopting domestic backend recently; Kunlunxin went from KernelGen operators (8/18) to vllm plugin backend (8/19) in just one day; the inference plugin line’s backend matrix now covers multiple vendors including Enflame, Iluvatar CoreX, Moore Threads, Ascend, Kunlunxin, and Xiwang Xinke. The training side (FlagScale CI expansion, Torch-FL AMP/profiler, FlagBLAS L2 completion) and the image side (build-infra app image unification) are advancing in parallel. It has been over 8 weeks since FlagOS 2.1 (6/24), with functionality continuing to accumulate, but this window still shows no signal of an official release.
III. Deep Dive on Member Units
Multi-vendor dynamics map within the window:
| Vendor | Chip/Backend | Dynamics within the window | Evidence |
|---|---|---|---|
| BAAI | — | build-infra leads the full-chain SUNRISE engineering (#441-#463); docs model list maintenance; no new model releases on the open-source side | build-infra commits / docs commits |
| SunRiseChip | SUNRISE | Window headline: FlagTree sunrise wheel production line formalized (#442) + vllm-sunrise 0.24.0 E2E on TANGRT 1.2.0 (#441) + decode hang fix and dual-version re-verification of 0.20.2/0.24.0 (#451/#459/#461) — moving from “backend support” into the “inference delivery” stage | build-infra #441-#463 |
| KunlunXin | KunlunXin | vllm-plugin-FL backend support merged for the first time (#268), Qwen3.6-35B-A3B/27B high-concurrency image-text verification passed | vllm-plugin-FL #268 |
| T-Head | Zhenwu PPU | Torch-FL PPU AMP integration completed (#140) | Torch-FL #140 |
| Moore Threads | MUSA | Torch-FL MUSA MUPTI profiler integration (#136); commercial side: another push toward a Hong Kong listing after nearing break-even | Torch-FL #136 / QQ News |
| Ascend (outside ecosystem) | Ascend | sglang-plugin-FL NPU e2e (#57); Torch-FL AMP/dtype extension (#135); FlagScale Ascend CI (#1250) | sglang-plugin-FL #57 / Torch-FL #135 / FlagScale #1250 |
| Iluvatar CoreX | Iluvatar | 2 Iluvatar-specific KernelGen operators in FlagGems-Experimental (special_modified_bessel_k0, index_copy_, #286) | FlagGems-Experimental #286 |
| MetaX | MetaX | FlagScale Metax CI (#1250); build-infra records metax flagtree training E2E (#435) and megatron-rl app environment (#443); commercial side: in-depth coverage of Tianji Intelligent Computing’s “Yunjian MetaX” | FlagScale #1250 / build-infra #435/#443 / gnews |
| Hygon | Hygon DCU | No new dynamics on the open-source side; commercial side: strategic cooperation with Tianji Data, joint innovation lab with Koal Software (8/19) | gnews multiple sources |
| Cambricon | Cambricon MLU | No new dynamics on the open-source side; commercial side: equity incentives for 124 employees unlocked and paid out (about 5.57 million yuan per person) | gnews multiple sources |
| Horizon Robotics | D-Robotics BPU | No dynamics on the open-source side; commercial side: Bosch MPC4 (Journey 6B) tens-of-millions-of-units-scale designation, Journey 6B lifetime designated volume exceeds 20 million units (8/19) | Guandian / Jiemian / DOIT |
| Enflame / Tsingmicro | Enflame GCU / TXDA | No open-source or news dynamics within the window | — |
Trend assessment: The chip landscape in this window is characterized by “deeply connected inference pipelines + continuously expanding backend matrix.” On one hand, SunRiseChip’s SUNRISE completed the most complete path for a domestic backend to connect to FlagOS (wheel production line → E2E → issue fixes), with KunlunXin following closely into the inference plugin line, bringing inference-side backend coverage to more than 6 vendors; on the other hand, the training side (FlagScale/FlagBLAS/Torch-FL) is advancing on three fronts: “multi-chip CI matrix + operator family completion + AMP/profiler capabilities.” On the commercial side, Horizon Robotics’ Journey 6B tens-of-millions-of-units designation, Hygon’s two consecutive ecosystem partnerships, and Moore Threads’ Hong Kong listing progress are the main highlights — software stack adaptation and commercial deployment are moving in step.
IV. Summary
- SUNRISE full inference pipeline connected (top story of this window): 12 build-infra commits completed the sunrise FlagTree wheel production line (#442, ubuntu 22.04 old-glibc base + LLVM 22 + sunriseTritonPlugin v0.6.0.4), vllm-sunrise 0.24.0 E2E (TANGRT 1.2.0, #441), decode hang fix and dual-version re-verification of 0.20.2/0.24.0 (#451/#459/#461) — the most complete engineering example of a domestic backend integrating with FlagOS.
- Kunlunxin enters the inference plugin line: vllm-plugin-FL #268 supports the Kunlunxin backend based on vLLM 0.20.2, with Qwen3.6-35B-A3B/27B high-concurrency text-and-image validation passing; only one day after its KernelGen operators were merged.
- Three parallel tracks on the training side: Torch-FL completed three chips on the same day (T-Head PPU AMP finished, Moore Threads MUSA MUPTI profiler, Ascend AMP extension); FlagScale added Ascend/MetaX build test CI; FlagBLAS filled out the Level-2 sparse/banded operator family within one day (SYMV/SPMV/GER/SYR/HER/SPR/HPR/SYR2/HER2/GBMV/HBMV/SBMV).
- Image and engineering mechanisms refined: build-infra unified the app image prefix to flagos-app and standardized tag naming (#457/#462/#463); added NVIDIA flash-attn source builds (#454) and cuda13.3 four-scenario dual-compiler verification records (#455); TransformerEngine-FL exposed the backend ops required for TP comm overlap (#95).
- Member companies active on the commercial front: Bosch’s MPC4 based on Journey 6B won a ten-million-unit-scale design win (Horizon, cumulative Journey 6B design wins surpassed 20 million units); Hygon reached two ecosystem partnerships with Tianji Data and Koal Software; Moore Threads is sprinting toward a Hong Kong listing; Cambricon’s incentive shares have unlocked and vested.
- Quiet on the news side: zero hits on component keywords, no valid HN results, no official GitHub Release; the FlagTensor 0.2.x and Megatron-LM-FL 0.2.0 series remain in preparation, with no version release signals yet.
Limitations: Commit counts and merge times come from commit search (the search index may lag slightly behind actual merges; individual commit times are based on the repos API pushed_at); the FlagBLAS series of commits mainly consists of batch merges from AriesMTR contributors; gnews redirect links cannot be scraped for full text, so news-side judgments rely on headlines and cross-verification from multiple sources; SUNRISE and Kunlunxin are not on the official list of org member units and are summarized based on backend ecosystem facts; the “Tianji Intelligent Computing” entry is an 8/19 in-depth report on the 7/24 launch event, not a new launch within this window.
Appendix: Complete Source List
| No. | Event | Source Link |
|---|---|---|
| 1 | build-infra SUNRISE wheel pipeline (#442) and vllm-sunrise 0.24.0 E2E (#441) | https://github.com/flagos-ai/build-infra/commits/main |
| 2 | build-infra SUNRISE decode hang fix and dual-version re-verification (#451/#459/#461) | https://github.com/flagos-ai/build-infra/commits/main |
| 3 | build-infra megatron app image mechanism (#443/#454/#455/#457/#462/#463) | https://github.com/flagos-ai/build-infra/commits/main |
| 4 | vllm-plugin-FL Kunlunxin backend (#268) | https://github.com/flagos-ai/vllm-plugin-FL/pull/268 |
| 5 | Torch-FL PPU AMP (#140), MUSA MUPTI (#136), Ascend AMP (#135), compile fixes (#132), and 6 others | https://github.com/flagos-ai/Torch-FL/commits/main |
| 6 | FlagScale Ascend/MetaX CI (#1250), TP overlap config (#1263), chat template (#1262) | https://github.com/flagos-ai/FlagScale/commits/main |
| 7 | FlagBLAS L2 operator family completion (AriesMTR PR #71-#81) | https://github.com/flagos-ai/FlagBLAS/commits/main |
| 8 | sglang-plugin-FL Ascend NPU e2e (#57) | https://github.com/flagos-ai/sglang-plugin-FL/pull/57 |
| 9 | FlagTree TLE layout anchor (#946), CI fixes (#1014) | https://github.com/flagos-ai/FlagTree/commits/main |
| 10 | FlagGems i0 optimization (#5592), vendor image update (#5628) | https://github.com/flagos-ai/FlagGems/commits/main |
| 11 | FlagGems-Experimental KernelGen Iluvatar operators (#286), split_with_sizes (#315) | https://github.com/flagos-ai/FlagGems-Experimental/commits/main |
| 12 | TransformerEngine-FL TP comm overlap ops (#95); FlagCX version metadata (#548) | https://github.com/flagos-ai/TransformerEngine-FL/commits/main |
| 13 | org repos overview (53 repos) | https://api.github.com/orgs/flagos-ai/repos?per_page=100&sort=updated |
| 14 | commit search (65 within window) | https://api.github.com/search/commits?q=org:flagos-ai+committer-date:%3E2026-08-19T02:19:00Z |
| 15 | Google News Chinese/English component queries (zero hits) | https://news.google.com/rss/search?q=FlagOS+when:7d |
| 16 | HN Algolia | https://hn.algolia.com/api/v1/search_by_date?query=FlagOS |
| 17 | Bosch’s 10-million-unit MPC4 design win based on Journey 6B (8/19) | https://www.guandian.cn/article/20260819/587064.html |
| 18 | Horizon Journey 6B design wins surpass 20 million units (8/19, DOIT) | https://www.doit.com.cn/ai/839586104311877.html |
| 19 | Tianji Data and Hygon strategic cooperation (8/19) | https://news.google.com/rss/articles/CBMiS0FVX3lxTE4tRS1NQS1id1ZmMXVrOTRWRk5mZEdfNkxqZFhiZGRHZUM3WjQwYm5XQTZ3T0pOMDFIbW5PWnh2ZGpza1NxdjhHYTV0NA?oc=5 |
| 20 | Koal Software and Hygon jointly establish innovation lab (8/19) | https://news.google.com/rss/articles/CBMiT0FVX3lxTE93VXhRYzdmVENMREhzRjZsOGZmYWtqNnJkd3dIazdkWWZNRDUteWFtWGpBNkFldWtENkFMVXZwbUxFY0kwbmpKTzJMT0I2YWs?oc=5 |
| 21 | Moore Threads returns for Hong Kong IPO “big test” (8/19, QQ News) | https://news.google.com/rss/articles/CBMiYkFVX3lxTE0wODJEUmIyZ1NhWjVSQ0tXQXJXWUJpTloxcTczSjNGUU1RMUYwUlVaMjF4SU96TUJHUkVFLXRhQnRWaWNWMXd4SDhzVnFCNFExYlNJV3FOUHR2VkVFTEdYdldR?oc=5 |
| 22 | Cloudpoint MetaX space-based intelligent computing report (8/19, Sohu) | https://news.google.com/rss/articles/CBMiiAFBVV95cUxObjR3US1fcW5iSXVPSnhGUnU2R1JsV1kxSTI3TXM2czhVbmZSX3hjcU5jSnNYNmM1cGdyZC1hOVJ2emlMUXl1QjZpNUxtY0pjd09HSGoyaS1IZ3lqVHNwbnpWak1kYTNvYzdBNk44aGxHbWRpN28zZEpqdmZhVDVRS2Qwb3NoTENH?oc=5 |
| 23 | Cambricon incentive unlock credited (8/19, Sohu) | https://news.google.com/rss/articles/CBMijAFBVV95cUxPQ2VVYW4wNUx3NXZEWjZMSHE2eXBkc0JkOVJOUGpFRDBpQU5qeklGZFgxb0pETGhzY3R6RTVnbFpleWEyQ0l1N0lwdmg3bGQyVGc4bHVOMERmOUp2QkdGZFpvN0w4YV80cTNMS2w3NGdJYTZVSGRaRnZoazVRdWlGX05RQnZib0lyZGZ3Rw?oc=5 |