The Beijing Academy of Artificial Intelligence (BAAI) has detailed FlagOS 2.2, the latest version of its open software stack for running AI models across different chip architectures. The release shipped on September 30 and was showcased at the 2026 AI Open Computing Conference in Beijing on October 17. BAAI describes the update as a shift from adapting models to many chips toward optimizing shared code across them.
The headline figure is scale. FlagOS 2.2 spans nine operator libraries, and public operators in FlagGems and related libraries grew from 564 to 1,454, an increase of 890. A new FlagGems-SGLang library adds 40 multi-chip operators covering attention, Mamba state updates, mixture-of-experts, LoRA, sampling and quantization. BAAI says that across more than a thousand operators on 10 kinds of AI hardware, overall operator performance rose by about 27%.
Much of the growth comes from automation. KernelGen, BAAI's automatic operator generator, now supports nine AI chips and produced 765 new operators this quarter, 416 of which passed accuracy checks on at least five domestic chips while reaching over 80% of native operator performance. FlagPrism, a unified debugging and profiling tool, links Triton source, compiler IR and device execution data, and supports Huawei Ascend, Iluvatar CoreX, Moore Threads, Nvidia and Enflame.
The release also adds FlagOS-Compressor for cross-chip weight conversion and quantization. BAAI says applying W8A8 to Qwen3.5-397B-A17B cut the model files from 751.41GB to 382.45GB, a 49.1% reduction, while scoring 84.85% on GPQA-Diamond. Combined with framework plugins for vLLM, Megatron-LM, SGLang and others, new models have been adapted to as many as 12 chip platforms within a day of release.
Case studies focus on older hardware. On Ascend 910B servers, BAAI reports that GLM-5.3-Flash logical serving throughput in an agentic replay test rose to 2.04 times its pre-optimization level, from 8.52k to 17.36k tokens per second. China Telecom's Xing4.0-29B-A4B model was adapted to 10 chip platforms, with time to first token down 19.87% in long-input tests compared with native operators.
BAAI argues the problem is timing: more than 80% of deployed AI compute is not the newest generation, while models change quickly. All figures come from BAAI's own tests, and it notes that real serving performance still depends on workload.
SEE ALSO: FlagOS 2.0 Expands Multi-Chip AI Ecosystem · Tsingway and BAAI Open-Source Open3D-PIMC