Today for AI
HOT RADAR
computeHEAT 11.7°

AI CURATED HOT STORY · 10/23/2026

BAAI Releases FlagOS 2.2 for Cross-Chip Operator Reuse and Performance Boost

1 reports archived1 independent sourcesupdated 10/23/2026, 20:00:37
Synthesis & Latest Updates
1 Sources Cross-Validated

BAAI released FlagOS 2.2, expanding its operator library to 1,454 entries and introducing specialized optimizations for SGLang and vLLM to enable cross-chip reuse across heterogeneous AI hardware. Benchmarks on Ascend 910B showed a 2x throughput increase for GLM-5.3-Flash and significant cost reductions, marking a shift in open computing from mere multi-chip support to continuous optimization for evolving models.

LATEST/BAAI released FlagOS 2.2, expanding its operator library to 1,454 entries and introducing specialized optimizations for SGLang and vLLM to enable cross-chip reuse across heterogeneous AI hardware. Benchmarks on Ascend 910B showed a 2x throughput increase for GLM-5.3-Flash and significant cost reductions, marking a shift in open computing from mere multi-chip support to continuous optimization for evolving models.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. InfoQ 中文T2·78 pts
    • FlagOS 2.2 expanded its operator library to 1,454 total operators (adding 890), covering general compute, inference, and scientific computing.
    • Cross-chip operator performance improved by ~27% across 10 heterogeneous AI chips; FlagGems-SGLang was released with 40 high-performance operators.
    • GLM-5.3-Flash achieved 2.04x logical service throughput on existing Ascend 910B hardware, reducing compute costs by 19.4%-27.6% compared to commercial API benchmarks.