Google 工程师教你怎么在手机上 21 分钟微调一个小模型，从 46% 干到 90%，比花 1500 美元买课还快。
在 AIE Europe 会议上，Google DeepMind 团队展示了一个完整微调流程：以 Gemma 270M 模型为起点，通过生成合成任务数据、LoRA 微调、int4 量化，最终部署到 Pixel 手机上。整个流程仅需 21 分钟，准确率从 46% 提升至 90%，推理速度达到 2000 tokens/s。该方法对比 1500 美元的线下 AI 训练营有显著效率和成本优势。
i guess this is a great time to note that @cormacb's talks and workshops were some of the most popul...
i guess this is a great time to note that @cormacb 's talks and workshops were some of the most popular @GoogleDeepMind sessions EVER! we're so glad to work with @osanseviero and @vadiamit in presenting AIE Europe, including a super well received keynote from @RaiaHadsell . Cormac returned with @benoitschilling for the World's Fair - videos launching today, see below 👇 h100envy @h100envy Google engineer explained how to fine-tune a tiny LLM from 46% to 90% accuracy on your phone in 21 minutes - better than $1500 on-device AI bootcamps. pick Gemma 270M -> generate synthetic task data -> fine-tune with LoRA -> quantize to int4 -> deploy to Pixel and hit 2000 tokens per second. That loop is how a 270M model beats a 70B one on your task, running fully offline in your pocket. Gemma 270M + synthetic data + LoRA + int4 quantization + on-device runtime - that's the stack. Watch and save it, then fine-tune your own tiny agent tonight. Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet 💬 2 🔄 1 ❤️ 9 👀 1663 📊 4 ⚡