Event ·Multi-source

Google将Gemma 4改造成扩散模型DiffusionGemma，训练预算不到10%

Google DeepMind将Gemma 4改造成扩散模型DiffusionGemma，训练成本不到原始预算的10%。该模型一次并行生成256个token，推理速度约每秒1500个token。在基准测试中，DiffusionGemma的质量仍落后于原始自回归模型，尤其在推理任务上差距明显。

Current read

Google DeepMind把Gemma 4改装成扩散模型DiffusionGemma，训练便宜十倍，生成快，但推理质量比原版差。适合想低成本做扩散模型的人看看。

4 sources68° AI heatLast updated 2026/8/9 10:01:26

Evidence

4 sources

Primary sourceDecoder

2026/8/9 10:01:26View source

Related source 1Google AI Developers

2026/8/10 21:10:57View source

Related source 2IT之家

2026/8/10 10:52:38View source

Related source 3The Rundown AI

2026/8/10 15:21:31View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story