JetBrains 工程师在 MacBook M5 Max 上用 Junie 本地运行 Qwen3.6-27B，代码与提示词均不上云，仅优化 agent harness 与推理配置（未改模型）即实现约 40% 的 prefill 提速。阿里云与 JetBrains 在 HumanX 上讨论了 agentic 开发落地生产所需的条件，核心观点是下一波 AI 增益将来自系统而非模型本身。
AI can generate code. But organizations still have to produce software.
Great conversation with @jetbrains at #HumanX on what it takes to make agentic development work in production. JetBrains engineers ran Qwen3.6-27B locally on a MacBook M5 Max with Junie — no code or prompts sent to the cloud. By optimizing the agent harness and inference config (not the model), they achieved ~40% faster prefill speed.
Key insight: the next big gain in AI won't come from the model alone — it will come from the system.
Open-weight models + cloud + local deployment = tools in the same toolbox.
The work can be delegated. Accountability cannot.
#AlibabaCloud #Qwen #AgenticAI
来源：@alibaba_cloud · x.com