DeepSeek unveils V4.1 Flash, 763B-parameter model that cuts memory needs
Chinese AI lab DeepSeek released V4.1 Flash, a 763-billion-parameter model with architectural improvements that cut key-value cache consumption to 13-25% of its predecessor, per The Register. The model can serve 4-8x more users in the same memory footprint, reducing GPU requirements for European companies deploying large language models.
Bottom line — DeepSeek V4.1 Flash reduces GPU memory required by roughly 25% versus its predecessor, per The Register.
Go deeper (5)
- The model introduces 'conditional memory module' using 196 billion n-gram parameters that can be offloaded to system RAM, so they don't need to be stored in GPU memory during inference, per The Register.
- Alibaba's Qwen 3.8-Flash-Next employs a similar n-gram approach from DeepSeek's January research, suggesting the technique will be widely adopted, per The Register.
- Despite being 2.5x larger than V4 Flash, the new model requires only 567 GB of GPU memory (FP8) versus 763 GB if all weights had to be held in GPU memory, per The Register.
- Google's Gemma team previously developed Per-Layer Embedding (PLE) tech for the same purpose of running LLMs on constrained devices like smartphones, per The Register.
- DeepSeek V4.1 Flash supports native multimodal visual understanding and a context window of 1M tokens, according to DeepSeek's announcement.