DiffusionGemma-26B-A4B-it — Google DeepMind 開源擴散式文字生成模型
- URL: https://www.threads.com/@dseditor/post/DZbgSxxGG_v
- Date Saved: 2026-06-11
- Source: Threads (@dseditor)
- Tags: creative-ai, ai-tools
Summary
dseditor 分享 Google DeepMind 開源的 DiffusionGemma-26B-A4B-it，引發討論。留言中 Meta AI bot 提供了詳細功能和硬體需求說明。
What is DiffusionGemma-26B-A4B-it?
Google DeepMind 開源的實驗性大模型，最大特色是使用「離散文字擴散」(discrete text diffusion) 而非傳統的逐 token 自回歸生成。
Key Features:
- Non-autoregressive decoding: 一次並行去噪一整塊文字，解碼速度快很多
- MoE architecture: 26B total parameters, only 8/128 experts activated per token (~3.8B active)
- Multimodal: 接受文字、圖片、影片輸入，輸出文字
- Supports system prompt + thinking mode
- Reasoning ability: 對標 Gemma 4
- License: Apache 2.0
Hardware Requirements:
- Minimum ~18GB VRAM (quantized)
- RTX 5090: 700+ tokens/s
- RTX 5060 Ti (16GB): can run Q4 quant with partial RAM offload (slower)
- Sweet spot: 20-30GB NVIDIA card for smooth performance
Deployment:
- HuggingFace weights (Apache 2.0)
- Supported: vLLM, Transformers, MLX
- NVIDIA NIM containerized microservice
- NeMo fine-tuning tutorials