China's Developer Ecosystem,
in Context.

China Dev Radar #003 — Verified AI Engineering, Linux Updates and Developer Tools

Six evidence-reviewed stories from November 2–6, 2026: Atrex supervised GPU optimization, Meituan LongCat agent-training engineering, Baidu ERNIE base-model weights, openEuler LTS maintenance, Tencent Hunyuan Hy4 visual computer-use announcement, and deepin desktop engineering research. Hy4 weights were forthcoming at announcement. The deepin report is a November-published retrospective of October work, not a new November release.

• Atrex Kernel Agent 0.4.0 introduces supervised GPU kernel optimization

  Atrex 0.4.0 introduces the Light workflow and modular provider architecture for GPU-kernel optimization with coding agents. An independent Supervisor controls correctness tests, source snapshots, same-allocation ABBA latency comparisons and candidate promotion. The release adds native CUDA manifests, multi-file sources and improved recovery, while migration from v0.3.0 requires workflow and workspace changes. Performance and token-savings claims remain publisher-reported.

• Meituan LongCat-2.5 details verifiable agent training and multimodal infrastructure

  Meituan's November 2 report describes a natively multimodal LongCat-2.5 model for long-horizon tasks, including complex codebase development. Its engineering account covers executable coding-task construction, fail-to-pass and pass-to-pass tests, asynchronous partial reinforcement-learning rollouts, and accelerator-aware long-context training and inference. Performance comparisons and throughput improvements are Meituan's own measurements, not independent benchmarks. This is distinct from the Meituan AI Coding article covered in Radar #002.

• Baidu releases ERNIE 5.5 Flash Base model weights and hybrid attention design

  Baidu released FP8 and BF16 pretrained base weights for ERNIE 5.5 Flash Base, a sparse MoE model with 280B total and about 14B active parameters. Its original architecture report describes compressed attention, selective full-context layers, and SGLang/vLLM-serving pathways. The 128K training context is distinct from configurations supporting up to 1M tokens in inference. This is a base checkpoint, not an instruction-tuned chat model, and the weights use a dedicated ERNIE model license. Benchmark comparisons are publisher-provided.