技術Google 官方開發者部落格 2026-09-23；Antigravity 為 Gemini CLI 官方後繼AI模型與API

Google 開發者部落格：Antigravity SDK 支援本地模型，Gemma 4 離線跑 agent

Introducing Support for Local AI Models in the Antigravity SDK

來源：Google Developers Blog・發布 2026-09-23・本站摘要分析（非全文轉載）

📌 一句話啟示：資料不能出公司的上班族，現在也能在自己的電腦跑 agent，零 API 費用。

摘要分析

Google 2026-09-23 宣布 Antigravity SDK（Gemini CLI 的後繼 agent 工具鏈）支援本地模型，首波是 Gemma 4 26B A4B 透過 LiteRT 執行，可離線、無 API 費用與速率限制、程式碼不出本機；也支援 Ollama、LM Studio、vLLM 等 OpenAI 相容伺服器，並示範雲端規劃、本地執行的混合模式。步驟：pip install google-antigravity litert-lm → litert-lm import 下載模型 → LiteRTAgentConfig 建 Agent 呼叫 agent.chat()。建議 24GB 以上記憶體。

python3 -m venv .venv && source .venv/bin/activate
pip install google-antigravity litert-lm
litert-lm import --from-huggingface-repo=litert-community/gemma-4-26B-A4B-it-litert-lm ... gemma4-26b
LiteRTAgentConfig(model_path=...).lightweight() → async with Agent(config) → agent.chat()
要寫檔案時傳 workspaces=[...] 與 policies=[policy.allow_all()]

On 2026-09-23 Google announced that the Antigravity SDK, the agent toolchain succeeding Gemini CLI, supports local models, starting with Gemma 4 26B A4B running on Google AI Edge's LiteRT. Developers can run agentic workflows offline with no API cost or rate limits while keeping code on their own machines. The SDK also works with OpenAI-compatible local servers such as Ollama, LM Studio and vLLM via LocalOpenAIAgentConfig, and the post describes a hybrid pattern where a cloud model plans and local models do the bulk of the work. Steps: create a virtual environment, pip install google-antigravity litert-lm, import Gemma 4 with litert-lm import from the Hugging Face repo, then build an Agent with LiteRTAgentConfig(model_path).lightweight() and stream agent.chat() tokens; pass workspaces and policies for file-writing tasks. A machine with more than 24GB of VRAM or unified memory is recommended.