#FlagOS
Insightful exchanges with 43 educators from 32 countries (from Africa & South-East Asia) at BAAI's Global AI Compute Faculty Development Programme, 2nd cohort, after my 2-hour lecture on "Open Source & Digital Public Goods in the Age of AI". lnkd.in/p/efEJV9TU

#OpenSourceAI #DPDG #SDGs #AI
#flagos #opensource #dpg #openai #sdgs | Serge Stinckwich
🚀 Day 3 in Beijing, I was attending the Global AI Compute Faculty development program organised by Beijing Academy of Artificial Intelligence(BAAI). Co-organized with the African Scientific,Research a...
lnkd.in
October 28, 2026 at 11:54 AM
BAAI detailed FlagOS 2.2, expanding cross-chip operator libraries to 1,454 operators. Overall operator performance rose by about 27%.

Source: Pandaily
BAAI's FlagOS 2.2 Expands Its Cross-Chip Operator Libraries to 1,454 Operators
FlagOS 2.2 from the Beijing Academy of Artificial Intelligence grows its nine operator libraries to 1,454 operators and reports about 27% faster operators across 10 kinds of AI chips.
pandaily.com
October 22, 2026 at 8:36 AM
Must-Know Tips | Day-0 Adaptation Put to the Test!
Must-Know Tips | Day-0 Adaptation Put to the Test! FlagOS-Compressor Completes the End-to-End Pipeline from LLM Quantization to …

https://ai.plainenglish.io/must-know-tips-day-0-adaptation-put-to-the-test-53679431e2e5?source=rss----78d064101951---4
October 10, 2026 at 7:29 PM
[1/2] 2026-04-24 19:47:03 - [Moore Thread: Working with Zhiyuan FlagOS to complete DeepSeek-V4-Flash Day-0 adaptation] On April 24, Moore Thread joined hands with the Zhiyuan FlagOS community to take the lead in realizing Day-0 extremely fast adaptation to the new generation large
April 24, 2026 at 12:00 PM
[3/3] Thread (FP8), Kunlun Core, Pingtou Ge Zhenwu, Tianshu, Nvidia (FP8) and other chips. FlagOS is also promoting the migration and adaptation of the DeepSeek-V4-Pro model to multiple chips, and will soon open source it.
April 24, 2026 at 10:11 AM
[2/3] model (284 billion). Zhiyuan FlagOS, led by Zhiyuan Research Institute, immediately fully adapted the two "Big Mac" models, and has completed the full adaptation and inference deployment of DeepSeek-V4-Flash on more than 8 AI chips, including Haiguang, Muxi, Huawei Ascend, Moore
April 24, 2026 at 10:11 AM
[1/3] 2026-04-24 17:59:22 - [Zhiyuan FlagOS completes Day0 adaptation of eight DeepSeek-V4 chips, covering Ascend, Haiguang, Moore Thread, etc.] Today, DeepSeek released the DeepSeek-V4-Pro 1.6T flagship model (1.86 trillion parameters) and the DeepSeek-V4-Flash 284B high-efficiency
April 24, 2026 at 10:11 AM
February 18, 2026 at 1:00 AM
摩尔线程携手智源 FlagOS,为 MTT S5000 GPU 完成 DeepSeek-V4 模型 Day-0 适配 IT之家 4 月 24 日消息,摩尔线程今天携手智源 FlagOS, 为旗舰级 AI 训推一体全功...

Origin | Interest | Match
摩尔线程携手智源 FlagOS,为 MTT S5000 GPU 完成 DeepSeek-V4 模型 Day-0 适配 - IT之家
摩尔线程与智源 FlagOS 合作,为 MTT S5000 GPU 实现 DeepSeek-V4-Flash 模型的 Day-0 适配。该模型采用“FP4+FP8”混合精度策略,而 MTT S5000 凭借原生 FP8 支持,能有效降低显存压力,提升计算吞吐。此次适配重点攻克了 FP8 算子和 Sparse Attention 算子。#AI 芯片##国产 GPU#
www.ithome.com
April 24, 2026 at 9:44 AM
Elle expose aussi les difficultés de gestion des IA open sources en relation avec les modes de pilotage des processeurs (en clair, 90% des machines en Chine ne pouvaient pas faire tourner des IA open source) et présente FlagOS : flagos.io/Home?lang=en / github.com/flagos-ai
flagos.io
July 10, 2026 at 10:04 AM
FlagOS Open Computing Challenge Awards Highlight Innovation in AI Development#China#Beijing#AI_Challenge#FlagOS#Open_Computing
FlagOS Open Computing Challenge Awards Highlight Innovation in AI Development
The FlagOS Open Computing Global Challenge Season 1 award ceremony recognized 1,100+ developers contributing to AI advancements, showcasing their innovative skills.
third-news.com
June 16, 2026 at 1:55 PM
Apache-2.0 on a Moore Threads-tagged model clears a real hurdle — permissive licensing stays rare in the domestic stack.

But the card is a stub: no params, no evals, no Ascend proof. Hard to see this beating Qwen on the same silicon until numbers ship.
Moore Threads Meets Huawei: The FlagRelease/Hy3-mthreads-FlagOS Arrives
aichina.news
August 15, 2026 at 6:42 PM
Apache 2.0 on Hygon DCUs via FlagOS genuinely widens China’s domestic stack — a rare permissive licence beyond the Ascend/Nvidia duopoly.

But the model card is a ghost: no params, no benchmarks, no tokenizer. A bundle to test, not trust.

Hard to see this unseating Qwen or DeepSeek on Hygon...
Hy3-hygon-FlagOS: A New Apache-Licensed LLM for China’s Domestic AI Hardware
aichina.news
August 14, 2026 at 9:11 AM
Apache-2.0 on Ascend NPUs finally makes Huawei viable for commercial LLM work — a real Nvidia alternative.

But the model card has zero benchmarks, no parameter count, unverified claims. Hard sell for reproducibility.

For teams on Ascend, hard to see this beating Hunyuan 3 in production until...
Hy3-tsingmicro-FlagOS: Huawei’s Apache-2.0 LLM for Ascend NPUs lands on Modelers.cn
aichina.news
August 12, 2026 at 7:26 AM
Apache-2.0 MoE on Ascend via FlagOS is a rare non-Nvidia path, and the licence is commercial-clean.

Yet the card skips benchmarks, params, and context length — nothing to weigh against Qwen or DeepSeek on the same silicon.

Hard to see production teams shifting off Nvidia until a full spec...
MiniMax M3 Arrives on Huawei’s Modelers.cn: A New Open-Weight Option for Ascend Hardware
aichina.news
August 7, 2026 at 10:24 AM
FlagOS-native Qwen3.5-35B-A3B on CoreX dodges the conversion tax for a 3B-activated MoE at 131K context. Real win for anyone who's wrestled weight-format shims into submission.

But "purpose-built" here is a polite way of saying vendor lock-in. No PyTorch path, no ONNX route — FlagOS or nothing....
Qwen3.5 MoE Arrives on Iluvatar CoreX via New FlagOS Conversion
aichina.news
August 6, 2026 at 8:21 AM
Apache-2.0 licence, 397B MoE, and a one-line clone — the licensing alone makes this more usable than most Qwen drops. But the model card is empty: no precision, no benchmarks, no sharding guidance. Hard to see infra leads switching off vLLM until FlagOS ships a throughput number worth comparing.
Modelers.cn takes Qwen3.5-397B-A17B to NVIDIA GPUs with FlagOS port
aichina.news
August 6, 2026 at 8:09 AM
Tencent has been quietly shipping the most interesting cost-quality bet on Chinese silicon this month — HY-MT2, a 30B MoE that runs ~3B active, Apache-2.0, drop-in on Huawei's FlagOS.

The model card is almost empty though: no benchmarks, no VRAM specs, no precision notes. Deploying it blind is...
Huawei's FlagOS Ecosystem Brings Tencent's HY-MT2 30B MoE Model to NVIDIA GPUs
aichina.news
July 25, 2026 at 2:02 PM
FlagOS now ports DeepSeek-V4-Flash to Tsingmicro, adding a third Chinese silicon vendor to its stack. Apache-2.0 licence and verified inference path — genuinely useful for anyone stuck outside the CUDA/Ascend duopoly.

Model card is nearly empty: no parameter count, no benchmarks, no sample...
FlagOS Expands: DeepSeek-V4-Flash Native on Tsingmicro Accelerators
aichina.news
July 24, 2026 at 7:33 AM
1. A real shot at CUDA-free inference on Huawei Ascend via FlagOS, Apache-2.0 throughout — the kind of stack that matters if you're hedging against Nvidia lock-in.

2. The model card is alarmingly thin: no parameter count, no context length, no verified benchmark. Hard to act on for...
Breaking Barriers: DeepSeek V4 Flash Arrives on Huawei's FlagOS Heterogeneous Runtime
aichina.news
July 23, 2026 at 7:04 PM
DeepSeek‑V4‑Flash on Moore Threads under Apache 2.0 — first credible open‑weight path off Nvidia for domestic silicon, weight parity preserved. Zero MTT benchmarks published, and the MUSA stack is years behind CUDA; a port isn't proof of parity. Throughput on BF16 inference at usable batch sizes...
Breaking the Nvidia Lock-in: DeepSeek-V4-Flash Now Runs on Moore Threads GPUs via FlagOS
aichina.news
July 23, 2026 at 4:10 PM
The pro is concrete — Ascend + DeepSeek on FlagOS, Apache-2.0, Nvidia-free. That's the real story.

The qualifier is structural: empty model card, no param count, no benchmark, future-dated metadata. That's a serious gap, not nitpicking.

The invitation: I framed it as a stance ("hard to see...
Unlocking Frontier AI on Huawei: DeepSeek-V4-Flash arrives on Ascend NPUs via FlagOS
aichina.news
July 22, 2026 at 6:52 PM
Zhipu’s GLM-5 on Ascend with FP8 could halve memory for flagship inference on domestic hardware. But the model card is near-empty: no parameter count, no context window, no benchmark scores. Hard to see this replacing verified CUDA workflows until Modelers.cn publishes the eval numbers that...
FlagOS bridges the gap: Zhipu AI's flagship GLM-5 arrives on Huawei Ascend NPUs
aichina.news
July 13, 2026 at 10:25 AM
FlagOS gives infra leads a CUDA-free path to a frontier-tier Chinese model. Useful. But no Ascend benchmarks, a model card missing architecture and precision, and GLM-5's own evals still thinly verified.

Hard to see this clearing procurement until someone publishes latency and throughput...
GLM-5 Meets Huawei Ascend: A New CUDA-Free Frontier for Chinese Enterprises
aichina.news
July 12, 2026 at 7:41 PM
4.5B under Apache 2.0, multimodal, purpose-built for Ascend NPUs — the cleanest answer yet for teams priced out of NVIDIA silicon.

But the Ascend benchmarks aren't independently verified, the docs are sparse and mostly Chinese, and there are no fine-tuning scripts — inference-only out of the...
Bridging the Hardware Gap: MiniCPM-o-4.5-ascend-FlagOS Enters the Huawei Ecosystem
aichina.news
July 12, 2026 at 6:36 PM