AINEWS
FlagOS is a BAAI led (Beijing Academy of AI) project that ported Alibaba's 2.4 trillion parameter Qwen3.8 to nine chip families, including eight Chinese designed accelerators, in under 24 hours.
A 2.4-trillion-parameter AI model from Alibaba's Qwen team is running on nine different chip families within hours of release, including eight Chinese-designed accelerators that, until recently, were considered too underpowered for frontier-scale work. The model is the vehicle. The story is the software layer that made it portable.
That layer is FlagOS, an open-source project led by the Beijing Academy of Artificial Intelligence (BAAI). According to Leiphone (雷锋网), a Chinese tech publication covering the announcement, FlagOS completed "Day-0" multi-chip adaptation for Qwen3.8-2.4T-A95B in under 24 hours. In practice, the model can be pulled and deployed on each chip family the same day it shipped, rather than weeks or months later when vendor-specific porting catches up.
The nine adapted chips are Pingtouge (Alibaba's own T-Head silicon), NVIDIA's accelerators, Moore Threads, Huawei Ascend, Muxi (MetaX), Kunlunxin, Hygon, Tsingmicro, and SuYuan. Eight of the nine are not Nvidia hardware. The ninth is.
The underlying model, Qwen3.8, is a mixture-of-experts system with 2.4 trillion total parameters but only 95 billion active at any single inference step. The active figure is the part that actually runs per token. The rest of the weights sit on chip and are consulted only as needed. It is the largest and most capable release in Alibaba's open Qwen series, and the first Qwen-Max-class model to ship as open weights rather than behind an API. The base model card is published on Hugging Face, with a separate FP8-quantized variant for hardware that supports that format.
The portability work is a translation layer that takes a model definition once and produces deployment images for each target chip, with precision alignment across BF16 (16-bit brain-float), FP8 (8-bit floating point), and INT8 (8-bit integer) formats. The FlagOS documentation describes it as a unified software stack. The plugin itself lives on GitHub as vllm-plugin-FL and sits on top of vLLM, SGLang, and Token Speed, the three most common open-source inference engines.
The second piece is a quantizer called FlagOS-Compressor. Per Leiphone, it converts a model's native FP8 or BF16 weights into INT8, a precision that, the report says, more than 80% of installed Chinese AI compute cannot run in FP8 natively. That figure is the publication's analytical framing and should be checked against an official FlagOS statement. The broader point holds, though. The installed base of AI accelerators in Chinese data centers is overwhelmingly BF16 and INT8, while frontier models are increasingly FP8-native. Without a software path between the two, the older chips get stranded.
The benchmark numbers in Leiphone's piece are self-reported by Alibaba and the FlagOS team. Qwen3.8 against the prior Qwen3.7-Max on coding-agent evals: Terminal Bench 2.1 rose from 74.5 to 86.6. SWE-bench Pro rose from 60.6 to 67.7, against Anthropic's Opus 4.8 at 69.2. Paper Bench rose from 64.8 to 93.0, with Opus 4.8 at 80.3. No independent reproduction is cited in the source, and the leaderboards themselves are not linked inline, so the deltas should be read as the vendor's claim, not as a settled result. The architectural change from Qwen3.5 and Qwen3.6 is also small, per the source. The engineering story is deployment at scale, not a model-architecture breakthrough.
Tencent Cloud's HAI community and 算力互联网, a Chinese compute-sharing network, are already shipping FlagOS-adapted model images in their container registries, according to Leiphone. A developer can pull a Qwen3.8 image and deploy a model-as-a-service endpoint on a domestic accelerator without writing chip-specific code.
The benchmark gains and the under-24-hour cycle are both vendor-defined. The leaderboards are not linked in the source, and "Day-0 adaptation" is a FlagOS metric, not an industry standard. The supported chip list also mixes one foreign accelerator, NVIDIA, with eight Chinese vendors, so the Day-0 story is genuinely about multi-vendor portability, not about Nvidia substitution in isolation.
The watch item is the next release. If FlagOS holds the cycle under 24 hours when the next frontier model lands, the portability layer starts looking like a competitive moat for whoever can run it fastest. The question for the rest of the industry is whether the same software problem gets solved once, or once per chip vendor.